chore: update demo runtime, tests, and agent docs

This commit is contained in:
2026-03-02 12:33:17 +08:00
parent 1f8f959ad7
commit cbb3284c13
14 changed files with 1491 additions and 236 deletions
+1
View File
@@ -147,4 +147,5 @@ dmypy.json
cython_debug/
ckpt/
output/
assets/*
+180 -73
View File
@@ -1,84 +1,191 @@
# PROJECT KNOWLEDGE BASE
# OpenGait Agent Guide
**Generated:** 2026-02-11T10:53:29Z
**Commit:** f754f6f
**Branch:** master
This file is for autonomous coding agents working in this repository.
Use it as the default playbook for commands, conventions, and safety checks.
## OVERVIEW
OpenGait is a research-grade, config-driven gait analysis framework centered on distributed PyTorch training/testing.
Core runtime lives in `opengait/`; `configs/` and `datasets/` are first-class operational surfaces, not just support folders.
## Scope and Ground Truth
## STRUCTURE
```text
OpenGait/
├── opengait/ # runtime package (train/test, model/data/eval pipelines)
├── configs/ # model- and dataset-specific experiment specs
├── datasets/ # preprocessing/rearrangement scripts + partitions
├── docs/ # user workflow docs
├── train.sh # launch patterns (DDP)
└── test.sh # eval launch patterns (DDP)
```
- Repository: `OpenGait`
- Runtime package: `opengait/`
- Primary entrypoint: `opengait/main.py`
- Package/runtime tool: `uv`
## WHERE TO LOOK
| Task | Location | Notes |
|------|----------|-------|
| Train/test entry | `opengait/main.py` | DDP init + config load + model dispatch |
| Model registration | `opengait/modeling/models/__init__.py` | dynamic class import/registration |
| Backbone/loss registration | `opengait/modeling/backbones/__init__.py`, `opengait/modeling/losses/__init__.py` | same dynamic pattern |
| Config merge behavior | `opengait/utils/common.py::config_loader` | merges into `configs/default.yaml` |
| Data loading contract | `opengait/data/dataset.py`, `opengait/data/collate_fn.py` | `.pkl` only, sequence sampling modes |
| Evaluation dispatch | `opengait/evaluation/evaluator.py` | dataset-specific eval routines |
| Dataset preprocessing | `datasets/pretreatment.py` + dataset subdirs | many standalone CLI tools |
Critical source-of-truth rule:
- `opengait/demo` is an implementation layer and may contain project-specific behavior.
- When asked to “refer to the paper” or verify methodology, use the paper and official citations as ground truth.
- Do not treat demo/runtime behavior as proof of paper method unless explicitly cited by the paper.
## CODE MAP
| Symbol / Module | Type | Location | Refs | Role |
|-----------------|------|----------|------|------|
| `config_loader` | function | `opengait/utils/common.py` | high | YAML merge + default overlay |
| `get_ddp_module` | function | `opengait/utils/common.py` | high | wraps modules with DDP passthrough |
| `BaseModel` | class | `opengait/modeling/base_model.py` | high | canonical train/test lifecycle |
| `LossAggregator` | class | `opengait/modeling/loss_aggregator.py` | medium | consumes `training_feat` contract |
| `DataSet` | class | `opengait/data/dataset.py` | high | dataset partition + sequence loading |
| `CollateFn` | class | `opengait/data/collate_fn.py` | high | fixed/unfixed/all sampling policy |
| `evaluate_*` funcs | functions | `opengait/evaluation/evaluator.py` | medium | metric/report orchestration |
| `models` package registry | dynamic module | `opengait/modeling/models/__init__.py` | high | config string → model class |
## Environment Setup
## CONVENTIONS
- Launch pattern is DDP-first (`python -m torch.distributed.launch ... opengait/main.py --cfgs ... --phase ...`).
- DDP Constraints: `world_size` must equal number of visible GPUs; test `evaluator_cfg.sampler.batch_size` must equal `world_size`.
- Model/loss/backbone discoverability is filesystem-driven via package-level dynamic imports.
- Experiment config semantics: custom YAML overlays `configs/default.yaml` (local key precedence).
- Outputs are keyed by config identity: `output/${dataset_name}/${model}/${save_name}`.
Install dependencies with uv:
## ANTI-PATTERNS (THIS PROJECT)
- Do not feed non-`.pkl` sequence files into runtime loaders (`opengait/data/dataset.py`).
- Do not violate sampler shape assumptions (`trainer_cfg.sampler.batch_size` is `[P, K]` for triplet regimes).
- Do not ignore DDP cleanup guidance; abnormal exits can leave zombie processes (`misc/clean_process.sh`).
- Do not add unregistered model/loss classes outside expected directories (`opengait/modeling/models`, `opengait/modeling/losses`).
## UNIQUE STYLES
- `datasets/` is intentionally script-heavy (rearrange/extract/pretreat), not a pure library package.
- Research model zoo is broad; many model files co-exist as first-class references.
- Recent repo trajectory includes scoliosis screening models (ScoNet lineage), not only person-ID gait benchmarks.
## COMMANDS
```bash
# install (uv)
uv sync --extra torch
# train (uv)
CUDA_VISIBLE_DEVICES=0,1 uv run python -m torch.distributed.launch --nproc_per_node=2 opengait/main.py --cfgs ./configs/baseline/baseline.yaml --phase train
# test (uv)
CUDA_VISIBLE_DEVICES=0,1 uv run python -m torch.distributed.launch --nproc_per_node=2 opengait/main.py --cfgs ./configs/baseline/baseline.yaml --phase test
# ScoNet 1-GPU eval
CUDA_VISIBLE_DEVICES=0 uv run python -m torch.distributed.launch --nproc_per_node=1 opengait/main.py --cfgs ./configs/sconet/sconet_scoliosis1k_local_eval_1gpu.yaml --phase test
# preprocess (generic)
python datasets/pretreatment.py --input_path <raw_or_rearranged> --output_path <pkl_root>
```
## NOTES
- LSP symbol map can be enabled via uv dev dependency `basedpyright`; `basedpyright` and `basedpyright-langserver` are available in `.venv` after `uv sync`.
- `train.sh` / `test.sh` are canonical launch examples across datasets/models.
- Academic-use-only restriction is stated in repository README.
Notes from `pyproject.toml`:
- Python requirement: `>=3.10`
- Dev tooling includes `pytest` and `basedpyright`
- Optional extras include `torch` and `parquet`
## Build / Run Commands
Train (DDP):
```bash
CUDA_VISIBLE_DEVICES=0,1 uv run python -m torch.distributed.launch \
--nproc_per_node=2 opengait/main.py \
--cfgs ./configs/baseline/baseline.yaml --phase train
```
Test (DDP):
```bash
CUDA_VISIBLE_DEVICES=0,1 uv run python -m torch.distributed.launch \
--nproc_per_node=2 opengait/main.py \
--cfgs ./configs/baseline/baseline.yaml --phase test
```
Single-GPU eval example:
```bash
CUDA_VISIBLE_DEVICES=0 uv run python -m torch.distributed.launch \
--nproc_per_node=1 opengait/main.py \
--cfgs ./configs/sconet/sconet_scoliosis1k_local_eval_1gpu.yaml --phase test
```
Demo CLI entry:
```bash
uv run python -m opengait.demo --help
```
## DDP Constraints (Important)
- `--nproc_per_node` must match visible GPU count in `CUDA_VISIBLE_DEVICES`.
- Test/evaluator sampling settings are strict and can fail if world size mismatches config.
- If interrupted DDP leaves stale processes:
```bash
sh misc/clean_process.sh
```
## Test Commands (especially single test)
Run all tests:
```bash
uv run pytest tests
```
Run one file:
```bash
uv run pytest tests/demo/test_pipeline.py -v
```
Run one test function:
```bash
uv run pytest tests/demo/test_pipeline.py::test_resolve_stride_modes -v
```
Run by keyword:
```bash
uv run pytest tests/demo/test_window.py -k "stride" -v
```
## Lint / Typecheck
Typecheck with basedpyright:
```bash
uv run basedpyright opengait tests
```
Project currently has no enforced formatter config in root tool files.
Follow existing local formatting and keep edits minimal.
## High-Value Paths
- `opengait/main.py` — runtime bootstrap
- `opengait/modeling/base_model.py` — model lifecycle contract
- `opengait/modeling/models/` — model zoo implementations
- `opengait/data/dataset.py` — dataset loading rules
- `opengait/data/collate_fn.py` — frame sampling behavior
- `opengait/evaluation/evaluator.py` — evaluation dispatch
- `configs/` — experiment definitions
- `datasets/` — preprocessing and partitions
## Code Style Guidelines
### Imports
- Keep ordering consistent: stdlib, third-party, local.
- Prefer explicit imports; avoid wildcard imports.
- Avoid introducing heavy imports in hot paths unless needed.
### Formatting
- Match surrounding file style (spacing, wrapping, structure).
- Avoid unrelated formatting churn.
- Keep diffs surgical.
### Types
- Add type annotations for new public APIs and non-trivial helpers.
- Reuse established typing style: `typing`, `numpy.typing`, `jaxtyping` where already used.
- Do not suppress type safety with blanket casts; keep unavoidable casts narrow.
### Naming
- `snake_case` for functions/variables
- `PascalCase` for classes
- `UPPER_SNAKE_CASE` for constants
- Preserve existing config key names and schema conventions
### Error Handling
- Raise explicit, actionable errors on invalid inputs.
- Fail fast for missing files, bad args, invalid shapes, and runtime preconditions.
- Never swallow exceptions silently.
- Preserve CLI error semantics (clear messages, non-zero exits).
### Logging
- Use module-level logger pattern already in codebase.
- Keep logs concise and operational.
- Avoid excessive per-frame logging in realtime/demo loops.
## Model and Config Contracts
- New models should conform to `BaseModel` expectations.
- Respect forward output dictionary contract used by loss/evaluator pipeline.
- Keep model registration/discovery patterns consistent with current package layout.
- Respect sampler semantics from config (`fixed_unordered`, `all_ordered`, etc.).
## Data Contracts
- Runtime data expects preprocessed `.pkl` sequence files.
- Partition JSON files are required for train/test split behavior.
- Do not mix modalities accidentally (silhouette / pose / pointcloud) across pipelines.
## Research-Verification Policy
When answering methodology questions:
- Prefer primary sources (paper PDF, official project docs, official code tied to publication).
- Quote/cite paper statements when concluding method behavior.
- If local implementation differs from paper, state divergence explicitly.
- For this repo specifically, remember: `opengait/demo` may differ from paper intent.
## Cursor / Copilot Rules Check
Checked these paths:
- `.cursor/rules/`
- `.cursorrules`
- `.github/copilot-instructions.md`
Current status: no Cursor/Copilot instruction files found.
## Agent Checklist Before Finishing
- Commands executed with `uv run ...` where applicable
- Targeted tests for changed files pass
- Typecheck is clean for modified code
- Behavior/documentation updated together for user-facing changes
- Paper-vs-implementation claims clearly separated when relevant
@@ -0,0 +1,101 @@
data_cfg:
dataset_name: Scoliosis1K
dataset_root: /mnt/public/data/Scoliosis1K/Scoliosis1K-sil-pkl
dataset_partition: /mnt/public/data/Scoliosis1K/Scoliosis1K_1116.json
num_workers: 1
remove_no_gallery: false
test_dataset_name: Scoliosis1K
evaluator_cfg:
enable_float16: true
restore_ckpt_strict: true
restore_hint: ./ckpt/ScoNet-20000.pt
save_name: ScoNet
eval_func: evaluate_scoliosis
sampler:
batch_shuffle: false
batch_size: 1
sample_type: all_ordered
frames_all_limit: 720
metric: euc
transform:
- type: BaseSilCuttingTransform
loss_cfg:
- loss_term_weight: 1.0
margin: 0.2
type: TripletLoss
log_prefix: triplet
- loss_term_weight: 1.0
scale: 16
type: CrossEntropyLoss
log_prefix: softmax
log_accuracy: true
model_cfg:
model: ScoNet
backbone_cfg:
type: ResNet9
block: BasicBlock
channels:
- 64
- 128
- 256
- 512
layers:
- 1
- 1
- 1
- 1
strides:
- 1
- 2
- 2
- 1
maxpool: false
SeparateFCs:
in_channels: 512
out_channels: 256
parts_num: 16
SeparateBNNecks:
class_num: 3
in_channels: 256
parts_num: 16
bin_num:
- 16
optimizer_cfg:
lr: 0.1
momentum: 0.9
solver: SGD
weight_decay: 0.0005
scheduler_cfg:
gamma: 0.1
milestones:
- 10000
- 14000
- 18000
scheduler: MultiStepLR
trainer_cfg:
enable_float16: true
fix_BN: false
with_test: false
log_iter: 100
restore_ckpt_strict: true
restore_hint: 0
save_iter: 20000
save_name: ScoNet
sync_BN: true
total_iter: 20000
sampler:
batch_shuffle: true
batch_size:
- 8
- 8
frames_num_fixed: 30
sample_type: fixed_unordered
type: TripletSampler
transform:
- type: BaseSilCuttingTransform
+144
View File
@@ -0,0 +1,144 @@
# Demo Window, Stride, and Sequence Behavior (ScoNet)
This note explains how the `opengait/demo` runtime feeds silhouettes into the neural network, what `stride` means, and when the sliding window is reset.
## Why sequence input (not single frame)
ScoNet-style inference is sequence-based.
- ScoNet / ScoNet-MT paper: https://arxiv.org/html/2407.05726v3
- DRF follow-up paper: https://arxiv.org/html/2509.00872v1
Both works use temporal information across walking frames rather than a single independent image.
### Direct quotes from the papers
From ScoNet / ScoNet-MT (MICCAI 2024, `2407.05726v3`):
> "For experiments, **30 frames were selected from each gait sequence as input**."
> (Section 4.1, Implementation Details)
From the same paper's dataset description:
> "Each sequence, containing approximately **300 frames at 15 frames per second**..."
> (Section 2.2, Data Collection and Preprocessing)
From DRF (MICCAI 2025, `2509.00872v1`):
DRF follows ScoNet-MT's sequence-level setup/architecture in its implementation details, and its PAV branch also aggregates across frames:
> "Sequence-Level PAV Refinement ... (2) **Temporal Aggregation**: For each metric, the mean of valid measurements across **all frames** is computed..."
> (Section 3.1, PAV: Discrete Clinical Prior)
## What papers say (and do not say) about stride
The papers define sequence-based inputs and temporal aggregation, but they do **not** define a deployment/runtime `stride` knob for online inference windows.
In other words:
- Paper gives the sequence framing (e.g., 30-frame inputs in ScoNet experiments).
- Demo `stride` is an engineering control for how often to run inference in streaming mode.
## What the demo feeds into the network
In `opengait/demo`, each inference uses the current silhouette buffer from `SilhouetteWindow`:
- Per-frame silhouette shape: `64 x 44`
- Tensor shape for inference: `[1, 1, window_size, 64, 44]`
- Default `window_size`: `30`
So by default, one prediction uses **30 silhouettes**.
## What is stride?
`stride` means the minimum frame distance between two consecutive classifications **after** the window is already full.
In this demo, the window is a true sliding buffer. It is **not** cleared after each inference. After inference, the pipeline only records the last classified frame and continues buffering new silhouettes.
- If `stride = 1`: classify at every new frame once ready
- If `stride = 30` (default): classify every 30 frames once ready
## Window mode shortcut (`--window-mode`)
To make window scheduling explicit, the demo CLI supports:
- `--window-mode manual` (default): use the exact `--stride` value
- `--window-mode sliding`: force `stride = 1` (max overlap)
- `--window-mode chunked`: force `stride = window` (no overlap)
This is only a shortcut for runtime behavior. It does not change ScoNet weights or architecture.
Examples:
- Sliding windows: `--window 30 --window-mode sliding` -> windows like `0-29, 1-30, 2-31, ...`
- Chunked windows: `--window 30 --window-mode chunked` -> windows like `0-29, 30-59, 60-89, ...`
- Manual stride: `--window 30 --stride 10 --window-mode manual` -> windows every 10 frames
Time interval between predictions is approximately:
`prediction_interval_seconds ~= stride / fps`
If `--target-fps` is set, use the emitted (downsampled) fps in this formula.
Examples:
- `stride=30`, `fps=15` -> about `2.0s`
- `stride=15`, `fps=30` -> about `0.5s`
First prediction latency is approximately:
`first_prediction_latency_seconds ~= window_size / fps`
assuming detections are continuous.
## Does the window clear when tracking target switches?
Yes. The window is reset in either case:
1. **Track ID changed** (new tracking target)
2. **Frame gap too large** (`frame_idx - last_frame > gap_threshold`)
Default `gap_threshold` in demo is `15` frames.
This prevents silhouettes from different people or long interrupted segments from being mixed into one inference window.
To be explicit:
- **Inference finished** -> window stays (sliding continues)
- **Track ID changed** -> window reset
- **Frame gap > gap_threshold** -> window reset
## Practical note about real-time detections
The window fills only when a valid silhouette is produced (i.e., person detection/segmentation succeeds). If detections are intermittent, the real-world time covered by one `window_size` can be longer than `window_size / fps`.
## Online vs offline behavior (important)
ScoNet's neural network does not hard-code a fixed frame count in the model graph. In OpenGait, frame count is controlled by sampling/runtime policy:
- Training config typically uses `frames_num_fixed: 30` with random fixed-frame sampling.
- Offline evaluation often uses `all_ordered` sequences (with `frames_all_limit` as a memory guard).
- Online demo uses the runtime window/stride scheduler.
So this does not mean the method only works offline. It means online performance depends on the latency/robustness trade-off you choose:
- Smaller windows / larger stride -> lower latency, potentially less stable predictions
- Larger windows / overlap -> smoother predictions, higher compute/latency
If you want behavior closest to ScoNet training assumptions, start from `--window 30` and tune stride (or `--window-mode`) for your deployment latency budget.
## Temporal downsampling (`--target-fps`)
Use `--target-fps` to normalize incoming frame cadence before silhouettes are pushed into the classification window.
- Default (`--target-fps 15`): timestamp-based pacing emits frames at approximately 15 FPS into the window
- Optional override (`--no-target-fps`): disable temporal downsampling and use all frames
Current default is `--target-fps 15` to align runtime cadence with ScoNet training assumptions.
For offline video sources, pacing uses video-time timestamps (`CAP_PROP_POS_MSEC`) when available, with an FPS-based synthetic timestamp fallback. This avoids coupling downsampling to processing throughput.
This is useful when camera FPS differs from training cadence. For example, with a 24 FPS camera:
- `--target-fps 15 --window 30` keeps model input near ~2.0 seconds of gait context (close to paper setup)
- `--stride` is interpreted in emitted-frame units after pacing
+53 -22
View File
@@ -3,8 +3,16 @@ from __future__ import annotations
import argparse
import logging
import sys
from typing import cast
from .pipeline import ScoliosisPipeline
from .pipeline import ScoliosisPipeline, WindowMode, resolve_stride
def _positive_float(value: str) -> float:
parsed = float(value)
if parsed <= 0:
raise argparse.ArgumentTypeError("target-fps must be positive")
return parsed
if __name__ == "__main__":
@@ -29,6 +37,24 @@ if __name__ == "__main__":
"--window", type=int, default=30, help="Window size for classification"
)
parser.add_argument("--stride", type=int, default=30, help="Stride for window")
parser.add_argument(
"--target-fps",
type=_positive_float,
default=15.0,
help="Target FPS for temporal downsampling before windowing",
)
parser.add_argument(
"--window-mode",
type=str,
choices=["manual", "sliding", "chunked"],
default="manual",
help="Window scheduling mode: manual uses --stride; sliding uses stride=1; chunked uses stride=window",
)
parser.add_argument(
"--no-target-fps",
action="store_true",
help="Disable temporal downsampling and use all frames",
)
parser.add_argument(
"--nats-url", type=str, default=None, help="NATS URL for result publishing"
)
@@ -88,27 +114,32 @@ if __name__ == "__main__":
source=args.source, checkpoint=args.checkpoint, config=args.config
)
# Build kwargs based on what ScoliosisPipeline accepts
pipeline_kwargs = {
"source": args.source,
"checkpoint": args.checkpoint,
"config": args.config,
"device": args.device,
"yolo_model": args.yolo_model,
"window": args.window,
"stride": args.stride,
"nats_url": args.nats_url,
"nats_subject": args.nats_subject,
"max_frames": args.max_frames,
"preprocess_only": args.preprocess_only,
"silhouette_export_path": args.silhouette_export_path,
"silhouette_export_format": args.silhouette_export_format,
"silhouette_visualize_dir": args.silhouette_visualize_dir,
"result_export_path": args.result_export_path,
"result_export_format": args.result_export_format,
"visualize": args.visualize,
}
pipeline = ScoliosisPipeline(**pipeline_kwargs)
effective_stride = resolve_stride(
window=cast(int, args.window),
stride=cast(int, args.stride),
window_mode=cast(WindowMode, args.window_mode),
)
pipeline = ScoliosisPipeline(
source=cast(str, args.source),
checkpoint=cast(str, args.checkpoint),
config=cast(str, args.config),
device=cast(str, args.device),
yolo_model=cast(str, args.yolo_model),
window=cast(int, args.window),
stride=effective_stride,
target_fps=(None if args.no_target_fps else cast(float, args.target_fps)),
nats_url=cast(str | None, args.nats_url),
nats_subject=cast(str, args.nats_subject),
max_frames=cast(int | None, args.max_frames),
preprocess_only=cast(bool, args.preprocess_only),
silhouette_export_path=cast(str | None, args.silhouette_export_path),
silhouette_export_format=cast(str, args.silhouette_export_format),
silhouette_visualize_dir=cast(str | None, args.silhouette_visualize_dir),
result_export_path=cast(str | None, args.result_export_path),
result_export_format=cast(str, args.result_export_format),
visualize=cast(bool, args.visualize),
)
raise SystemExit(pipeline.run())
except ValueError as err:
print(f"Error: {err}", file=sys.stderr)
+17 -2
View File
@@ -18,6 +18,7 @@ logger = logging.getLogger(__name__)
# Type alias for frame stream: (frame_array, metadata_dict)
FrameStream = Iterable[tuple[np.ndarray, dict[str, object]]]
# Protocol for cv-mmap metadata (needed at runtime for nested function annotation)
class _FrameMetadata(Protocol):
frame_count: int
@@ -58,6 +59,13 @@ def opencv_source(
if not cap.isOpened():
raise RuntimeError(f"Failed to open video source: {path}")
is_file_source = isinstance(path, str)
source_fps = float(cap.get(cv2.CAP_PROP_FPS)) if is_file_source else 0.0
fps_valid = source_fps > 0.0 and np.isfinite(source_fps)
fallback_fps = source_fps if fps_valid else 30.0
fallback_interval_ns = int(1_000_000_000 / fallback_fps)
start_ns = time.monotonic_ns()
frame_idx = 0
try:
while max_frames is None or frame_idx < max_frames:
@@ -66,7 +74,13 @@ def opencv_source(
# End of stream
break
# Get timestamp if available (some backends support this)
if is_file_source:
pos_msec = float(cap.get(cv2.CAP_PROP_POS_MSEC))
if np.isfinite(pos_msec) and pos_msec > 0.0:
timestamp_ns = start_ns + int(pos_msec * 1_000_000)
else:
timestamp_ns = start_ns + frame_idx * fallback_interval_ns
else:
timestamp_ns = time.monotonic_ns()
metadata: dict[str, object] = {
@@ -74,6 +88,8 @@ def opencv_source(
"timestamp_ns": timestamp_ns,
"source": path,
}
if fps_valid:
metadata["source_fps"] = source_fps
yield frame, metadata
frame_idx += 1
@@ -118,7 +134,6 @@ def cvmmap_source(
# Import cvmmap only when function is called
# Use try/except for runtime import check
try:
from cvmmap import CvMmapClient as _CvMmapClientReal # pyright: ignore[reportMissingTypeStubs]
except ImportError as e:
raise ImportError(
+105 -5
View File
@@ -5,7 +5,7 @@ from contextlib import suppress
import logging
from pathlib import Path
import time
from typing import TYPE_CHECKING, Protocol, cast
from typing import TYPE_CHECKING, Literal, Protocol, TypeAlias, cast
from beartype import beartype
import click
@@ -31,6 +31,16 @@ JaxtypedDecorator = Callable[[Callable[..., object]], Callable[..., object]]
JaxtypedFactory = Callable[..., JaxtypedDecorator]
jaxtyped = cast(JaxtypedFactory, jaxtyping.jaxtyped)
WindowMode: TypeAlias = Literal["manual", "sliding", "chunked"]
def resolve_stride(window: int, stride: int, window_mode: WindowMode) -> int:
if window_mode == "manual":
return stride
if window_mode == "sliding":
return 1
return window
class _BoxesLike(Protocol):
@property
@@ -65,6 +75,27 @@ class _TrackCallable(Protocol):
) -> object: ...
class _FramePacer:
_interval_ns: int
_next_emit_ns: int | None
def __init__(self, target_fps: float) -> None:
if target_fps <= 0:
raise ValueError(f"target_fps must be positive, got {target_fps}")
self._interval_ns = int(1_000_000_000 / target_fps)
self._next_emit_ns = None
def should_emit(self, timestamp_ns: int) -> bool:
if self._next_emit_ns is None:
self._next_emit_ns = timestamp_ns + self._interval_ns
return True
if timestamp_ns >= self._next_emit_ns:
while self._next_emit_ns <= timestamp_ns:
self._next_emit_ns += self._interval_ns
return True
return False
class ScoliosisPipeline:
_detector: object
_source: FrameStream
@@ -83,6 +114,7 @@ class ScoliosisPipeline:
_result_buffer: list[DemoResult]
_visualizer: OpenCVVisualizer | None
_last_viz_payload: dict[str, object] | None
_frame_pacer: _FramePacer | None
def __init__(
self,
@@ -104,6 +136,7 @@ class ScoliosisPipeline:
result_export_path: str | None = None,
result_export_format: str = "json",
visualize: bool = False,
target_fps: float | None = 15.0,
) -> None:
self._detector = YOLO(yolo_model)
self._source = create_source(source, max_frames=max_frames)
@@ -140,6 +173,7 @@ class ScoliosisPipeline:
else:
self._visualizer = None
self._last_viz_payload = None
self._frame_pacer = _FramePacer(target_fps) if target_fps is not None else None
@staticmethod
def _extract_int(meta: dict[str, object], key: str, fallback: int) -> int:
@@ -177,6 +211,7 @@ class ScoliosisPipeline:
Float[ndarray, "64 44"],
UInt8[ndarray, "h w"],
BBoxXYXY,
BBoxXYXY,
int,
]
| None
@@ -189,7 +224,7 @@ class ScoliosisPipeline:
mask_to_silhouette(self._to_mask_u8(mask_raw), bbox_mask),
)
if silhouette is not None:
return silhouette, mask_raw, bbox_frame, int(track_id)
return silhouette, mask_raw, bbox_frame, bbox_mask, int(track_id)
fallback = cast(
tuple[UInt8[ndarray, "h w"], BBoxXYXY] | None,
@@ -231,7 +266,7 @@ class ScoliosisPipeline:
# Fallback: use mask-space bbox if orig_shape unavailable
bbox_frame = bbox_mask
# For fallback case, mask_raw is the same as mask_u8
return silhouette, mask_u8, bbox_frame, 0
return silhouette, mask_u8, bbox_frame, bbox_mask, 0
@jaxtyped(typechecker=beartype)
def process_frame(
@@ -262,7 +297,7 @@ class ScoliosisPipeline:
if selected is None:
return None
silhouette, mask_raw, bbox, track_id = selected
silhouette, mask_raw, bbox, bbox_mask, track_id = selected
# Store silhouette for export if in preprocess-only mode or if export requested
if self._silhouette_export_path is not None or self._preprocess_only:
@@ -284,20 +319,39 @@ class ScoliosisPipeline:
return {
"mask_raw": mask_raw,
"bbox": bbox,
"bbox_mask": bbox_mask,
"silhouette": silhouette,
"segmentation_input": None,
"track_id": track_id,
"label": None,
"confidence": None,
}
if self._frame_pacer is not None and not self._frame_pacer.should_emit(
timestamp_ns
):
return {
"mask_raw": mask_raw,
"bbox": bbox,
"bbox_mask": bbox_mask,
"silhouette": silhouette,
"segmentation_input": self._window.buffered_silhouettes,
"track_id": track_id,
"label": None,
"confidence": None,
}
self._window.push(silhouette, frame_idx=frame_idx, track_id=track_id)
segmentation_input = self._window.buffered_silhouettes
if not self._window.should_classify():
# Return visualization payload even when not classifying yet
return {
"mask_raw": mask_raw,
"bbox": bbox,
"bbox_mask": bbox_mask,
"silhouette": silhouette,
"segmentation_input": segmentation_input,
"track_id": track_id,
"label": None,
"confidence": None,
@@ -330,7 +384,9 @@ class ScoliosisPipeline:
"result": result,
"mask_raw": mask_raw,
"bbox": bbox,
"bbox_mask": bbox_mask,
"silhouette": silhouette,
"segmentation_input": segmentation_input,
"track_id": track_id,
"label": label,
"confidence": confidence,
@@ -400,7 +456,9 @@ class ScoliosisPipeline:
viz_dict = cast(dict[str, object], viz_data)
mask_raw_obj = viz_dict.get("mask_raw")
bbox_obj = viz_dict.get("bbox")
bbox_mask_obj = viz_dict.get("bbox_mask")
silhouette_obj = viz_dict.get("silhouette")
segmentation_input_obj = viz_dict.get("segmentation_input")
track_id_val = viz_dict.get("track_id", 0)
track_id = track_id_val if isinstance(track_id_val, int) else 0
label_obj = viz_dict.get("label")
@@ -409,24 +467,33 @@ class ScoliosisPipeline:
# Cast extracted values to expected types
mask_raw = cast(NDArray[np.uint8] | None, mask_raw_obj)
bbox = cast(BBoxXYXY | None, bbox_obj)
bbox_mask = cast(BBoxXYXY | None, bbox_mask_obj)
silhouette = cast(NDArray[np.float32] | None, silhouette_obj)
segmentation_input = cast(
NDArray[np.float32] | None,
segmentation_input_obj,
)
label = cast(str | None, label_obj)
confidence = cast(float | None, confidence_obj)
else:
# No detection and no cache - use default values
mask_raw = None
bbox = None
bbox_mask = None
track_id = 0
silhouette = None
segmentation_input = None
label = None
confidence = None
keep_running = self._visualizer.update(
frame_u8,
bbox,
bbox_mask,
track_id,
mask_raw,
silhouette,
segmentation_input,
label,
confidence,
ema_fps,
@@ -671,6 +738,23 @@ def validate_runtime_inputs(source: str, checkpoint: str, config: str) -> None:
)
@click.option("--window", type=click.IntRange(min=1), default=30, show_default=True)
@click.option("--stride", type=click.IntRange(min=1), default=30, show_default=True)
@click.option(
"--window-mode",
type=click.Choice(["manual", "sliding", "chunked"], case_sensitive=False),
default="manual",
show_default=True,
help=(
"Window scheduling mode: manual uses --stride; "
"sliding forces stride=1; chunked forces stride=window"
),
)
@click.option(
"--target-fps",
type=click.FloatRange(min=0.1),
default=15.0,
show_default=True,
)
@click.option("--no-target-fps", is_flag=True, default=False)
@click.option("--nats-url", type=str, default=None)
@click.option(
"--nats-subject",
@@ -725,6 +809,9 @@ def main(
yolo_model: str,
window: int,
stride: int,
window_mode: str,
target_fps: float | None,
no_target_fps: bool,
nats_url: str | None,
nats_subject: str,
max_frames: int | None,
@@ -748,6 +835,18 @@ def main(
try:
validate_runtime_inputs(source=source, checkpoint=checkpoint, config=config)
effective_stride = resolve_stride(
window=window,
stride=stride,
window_mode=cast(WindowMode, window_mode.lower()),
)
if effective_stride != stride:
logger.info(
"window_mode=%s overrides stride=%d -> effective_stride=%d",
window_mode,
stride,
effective_stride,
)
pipeline = ScoliosisPipeline(
source=source,
checkpoint=checkpoint,
@@ -755,7 +854,8 @@ def main(
device=device,
yolo_model=yolo_model,
window=window,
stride=stride,
stride=effective_stride,
target_fps=None if no_target_fps else target_fps,
nats_url=nats_url,
nats_subject=nats_subject,
max_frames=max_frames,
+3 -1
View File
@@ -23,8 +23,10 @@ jaxtyped = cast(JaxtypedFactory, jaxtyping.jaxtyped)
UInt8Array = NDArray[np.uint8]
Float32Array = NDArray[np.float32]
#: Bounding box in XYXY format: (x1, y1, x2, y2) where (x1,y1) is top-left and (x2,y2) is bottom-right.
BBoxXYXY = tuple[int, int, int, int]
"""
Bounding box in XYXY format: (x1, y1, x2, y2) where (x1,y1) is top-left and (x2,y2) is bottom-right.
"""
def _read_attr(container: object, key: str) -> object | None:
+247 -123
View File
@@ -20,7 +20,9 @@ logger = logging.getLogger(__name__)
# Window names
MAIN_WINDOW = "Scoliosis Detection"
SEG_WINDOW = "Segmentation"
SEG_WINDOW = "Normalized Silhouette"
RAW_WINDOW = "Raw Mask"
WINDOW_SEG_INPUT = "Segmentation Input"
# Silhouette dimensions (from preprocess.py)
SIL_HEIGHT = 64
@@ -29,43 +31,45 @@ SIL_WIDTH = 44
# Display dimensions for upscaled silhouette
DISPLAY_HEIGHT = 256
DISPLAY_WIDTH = 176
RAW_STATS_PAD = 54
MODE_LABEL_PAD = 26
# Colors (BGR)
COLOR_GREEN = (0, 255, 0)
COLOR_WHITE = (255, 255, 255)
COLOR_BLACK = (0, 0, 0)
COLOR_DARK_GRAY = (56, 56, 56)
COLOR_RED = (0, 0, 255)
COLOR_YELLOW = (0, 255, 255)
# Mode labels
MODE_LABELS = ["Both", "Raw Mask", "Normalized"]
# Type alias for image arrays (NDArray or cv2.Mat)
ImageArray = NDArray[np.uint8]
class OpenCVVisualizer:
"""Real-time visualizer for gait analysis demo.
Displays two windows:
- Main stream: Original frame with bounding box and metadata overlay
- Segmentation: Raw mask, normalized silhouette, or side-by-side view
Supports interactive mode switching via keyboard.
"""
def __init__(self) -> None:
"""Initialize visualizer with default mask mode."""
self.mask_mode: int = 0 # 0: Both, 1: Raw, 2: Normalized
self.show_raw_window: bool = False
self.show_raw_debug: bool = False
self._windows_created: bool = False
self._raw_window_created: bool = False
def _ensure_windows(self) -> None:
"""Create OpenCV windows if not already created."""
if not self._windows_created:
cv2.namedWindow(MAIN_WINDOW, cv2.WINDOW_NORMAL)
cv2.namedWindow(SEG_WINDOW, cv2.WINDOW_NORMAL)
cv2.namedWindow(WINDOW_SEG_INPUT, cv2.WINDOW_NORMAL)
self._windows_created = True
def _ensure_raw_window(self) -> None:
if not self._raw_window_created:
cv2.namedWindow(RAW_WINDOW, cv2.WINDOW_NORMAL)
self._raw_window_created = True
def _hide_raw_window(self) -> None:
if self._raw_window_created:
cv2.destroyWindow(RAW_WINDOW)
self._raw_window_created = False
def _draw_bbox(
self,
frame: ImageArray,
@@ -215,33 +219,181 @@ class OpenCVVisualizer:
return upscaled
def _normalize_mask_for_display(self, mask: NDArray[np.generic]) -> ImageArray:
mask_array = np.asarray(mask)
if mask_array.dtype == np.bool_:
bool_scaled = np.where(mask_array, np.uint8(255), np.uint8(0)).astype(
np.uint8
)
return cast(ImageArray, bool_scaled)
if mask_array.dtype == np.uint8:
mask_array = cast(ImageArray, mask_array)
max_u8 = int(np.max(mask_array)) if mask_array.size > 0 else 0
if max_u8 <= 1:
scaled_u8 = np.where(mask_array > 0, np.uint8(255), np.uint8(0)).astype(
np.uint8
)
return cast(ImageArray, scaled_u8)
return cast(ImageArray, mask_array)
if np.issubdtype(mask_array.dtype, np.integer):
max_int = float(np.max(mask_array)) if mask_array.size > 0 else 0.0
if max_int <= 1.0:
return cast(
ImageArray, (mask_array.astype(np.float32) * 255.0).astype(np.uint8)
)
clipped = np.clip(mask_array, 0, 255).astype(np.uint8)
return cast(ImageArray, clipped)
mask_float = np.asarray(mask_array, dtype=np.float32)
max_val = float(np.max(mask_float)) if mask_float.size > 0 else 0.0
if max_val <= 0.0:
return np.zeros(mask_float.shape, dtype=np.uint8)
normalized = np.clip((mask_float / max_val) * 255.0, 0.0, 255.0).astype(
np.uint8
)
return cast(ImageArray, normalized)
def _draw_raw_stats(self, image: ImageArray, mask_raw: ImageArray | None) -> None:
if mask_raw is None:
return
mask = np.asarray(mask_raw)
if mask.size == 0:
return
stats = [
f"raw: {mask.dtype}",
f"min/max: {float(mask.min()):.3f}/{float(mask.max()):.3f}",
f"nnz: {int(np.count_nonzero(mask))}",
]
font = cv2.FONT_HERSHEY_SIMPLEX
font_scale = 0.45
thickness = 1
line_h = 18
x0 = 8
y0 = 20
for i, txt in enumerate(stats):
y = y0 + i * line_h
(tw, th), _ = cv2.getTextSize(txt, font, font_scale, thickness)
_ = cv2.rectangle(
image, (x0 - 4, y - th - 4), (x0 + tw + 4, y + 4), COLOR_BLACK, -1
)
_ = cv2.putText(
image, txt, (x0, y), font, font_scale, COLOR_YELLOW, thickness
)
def _prepare_segmentation_view(
self,
mask_raw: ImageArray | None,
silhouette: NDArray[np.float32] | None,
bbox: BBoxXYXY | None,
) -> ImageArray:
"""Prepare segmentation window content based on current mode.
Args:
mask_raw: Raw binary mask (H, W) uint8 or None
silhouette: Normalized silhouette (64, 44) float32 or None
Returns:
Displayable image (H, W, 3) uint8
"""
if self.mask_mode == 0:
# Mode 0: Both (side by side)
return self._prepare_both_view(mask_raw, silhouette)
elif self.mask_mode == 1:
# Mode 1: Raw mask only
return self._prepare_raw_view(mask_raw)
else:
# Mode 2: Normalized silhouette only
_ = mask_raw
_ = bbox
return self._prepare_normalized_view(silhouette)
def _fit_gray_to_display(
self,
gray: ImageArray,
out_h: int = DISPLAY_HEIGHT,
out_w: int = DISPLAY_WIDTH,
) -> ImageArray:
src_h, src_w = gray.shape[:2]
if src_h <= 0 or src_w <= 0:
return np.zeros((out_h, out_w), dtype=np.uint8)
scale = min(out_w / src_w, out_h / src_h)
new_w = max(1, int(round(src_w * scale)))
new_h = max(1, int(round(src_h * scale)))
resized = cast(
ImageArray,
cv2.resize(gray, (new_w, new_h), interpolation=cv2.INTER_NEAREST),
)
canvas = np.zeros((out_h, out_w), dtype=np.uint8)
x0 = (out_w - new_w) // 2
y0 = (out_h - new_h) // 2
canvas[y0 : y0 + new_h, x0 : x0 + new_w] = resized
return cast(ImageArray, canvas)
def _crop_mask_to_bbox(
self,
mask_gray: ImageArray,
bbox: BBoxXYXY | None,
) -> ImageArray:
if bbox is None:
return mask_gray
h, w = mask_gray.shape[:2]
x1, y1, x2, y2 = bbox
x1c = max(0, min(w, int(x1)))
x2c = max(0, min(w, int(x2)))
y1c = max(0, min(h, int(y1)))
y2c = max(0, min(h, int(y2)))
if x2c <= x1c or y2c <= y1c:
return mask_gray
cropped = mask_gray[y1c:y2c, x1c:x2c]
if cropped.size == 0:
return mask_gray
return cast(ImageArray, cropped)
def _prepare_segmentation_input_view(
self,
silhouettes: NDArray[np.float32] | None,
) -> ImageArray:
if silhouettes is None or silhouettes.size == 0:
placeholder = np.zeros((DISPLAY_HEIGHT, DISPLAY_WIDTH, 3), dtype=np.uint8)
self._draw_mode_indicator(placeholder, "Input Silhouettes (No Data)")
return placeholder
n_frames = int(silhouettes.shape[0])
tiles_per_row = int(np.ceil(np.sqrt(n_frames)))
rows = int(np.ceil(n_frames / tiles_per_row))
tile_h = DISPLAY_HEIGHT
tile_w = DISPLAY_WIDTH
grid = np.zeros((rows * tile_h, tiles_per_row * tile_w), dtype=np.uint8)
for idx in range(n_frames):
sil = silhouettes[idx]
tile = self._upscale_silhouette(sil)
r = idx // tiles_per_row
c = idx % tiles_per_row
y0, y1 = r * tile_h, (r + 1) * tile_h
x0, x1 = c * tile_w, (c + 1) * tile_w
grid[y0:y1, x0:x1] = tile
grid_bgr = cast(ImageArray, cv2.cvtColor(grid, cv2.COLOR_GRAY2BGR))
for idx in range(n_frames):
r = idx // tiles_per_row
c = idx % tiles_per_row
y0 = r * tile_h
x0 = c * tile_w
cv2.putText(
grid_bgr,
str(idx),
(x0 + 8, y0 + 22),
cv2.FONT_HERSHEY_SIMPLEX,
0.6,
(0, 255, 255),
2,
cv2.LINE_AA,
)
return grid_bgr
def _prepare_raw_view(
self,
mask_raw: ImageArray | None,
bbox: BBoxXYXY | None = None,
) -> ImageArray:
"""Prepare raw mask view.
@@ -261,20 +413,23 @@ class OpenCVVisualizer:
if len(mask_raw.shape) == 3:
mask_gray = cast(ImageArray, cv2.cvtColor(mask_raw, cv2.COLOR_BGR2GRAY))
else:
mask_gray = mask_raw
mask_gray = cast(ImageArray, mask_raw)
# Resize to display size
mask_resized = cast(
ImageArray,
cv2.resize(
mask_gray,
(DISPLAY_WIDTH, DISPLAY_HEIGHT),
interpolation=cv2.INTER_NEAREST,
),
mask_gray = self._normalize_mask_for_display(mask_gray)
mask_gray = self._crop_mask_to_bbox(mask_gray, bbox)
debug_pad = RAW_STATS_PAD if self.show_raw_debug else 0
content_h = max(1, DISPLAY_HEIGHT - debug_pad - MODE_LABEL_PAD)
mask_resized = self._fit_gray_to_display(
mask_gray, out_h=content_h, out_w=DISPLAY_WIDTH
)
full_mask = np.zeros((DISPLAY_HEIGHT, DISPLAY_WIDTH), dtype=np.uint8)
full_mask[debug_pad : debug_pad + content_h, :] = mask_resized
# Convert to BGR for display
mask_bgr = cast(ImageArray, cv2.cvtColor(mask_resized, cv2.COLOR_GRAY2BGR))
mask_bgr = cast(ImageArray, cv2.cvtColor(full_mask, cv2.COLOR_GRAY2BGR))
if self.show_raw_debug:
self._draw_raw_stats(mask_bgr, mask_raw)
self._draw_mode_indicator(mask_bgr, "Raw Mask")
return mask_bgr
@@ -299,80 +454,21 @@ class OpenCVVisualizer:
# Upscale and convert
upscaled = self._upscale_silhouette(silhouette)
sil_bgr = cast(ImageArray, cv2.cvtColor(upscaled, cv2.COLOR_GRAY2BGR))
content_h = max(1, DISPLAY_HEIGHT - MODE_LABEL_PAD)
sil_compact = self._fit_gray_to_display(
upscaled, out_h=content_h, out_w=DISPLAY_WIDTH
)
sil_canvas = np.zeros((DISPLAY_HEIGHT, DISPLAY_WIDTH), dtype=np.uint8)
sil_canvas[:content_h, :] = sil_compact
sil_bgr = cast(ImageArray, cv2.cvtColor(sil_canvas, cv2.COLOR_GRAY2BGR))
self._draw_mode_indicator(sil_bgr, "Normalized")
return sil_bgr
def _prepare_both_view(
self,
mask_raw: ImageArray | None,
silhouette: NDArray[np.float32] | None,
) -> ImageArray:
"""Prepare side-by-side view of both masks.
Args:
mask_raw: Raw binary mask or None
silhouette: Normalized silhouette or None
Returns:
Displayable side-by-side image
"""
# Prepare individual views without mode indicators (will be drawn on combined)
# Raw view preparation (without indicator)
if mask_raw is None:
raw_gray = np.zeros((DISPLAY_HEIGHT, DISPLAY_WIDTH), dtype=np.uint8)
else:
if len(mask_raw.shape) == 3:
mask_gray = cast(ImageArray, cv2.cvtColor(mask_raw, cv2.COLOR_BGR2GRAY))
else:
mask_gray = mask_raw
# Normalize to uint8 [0,255] for display (handles both float [0,1] and uint8 inputs)
if mask_gray.dtype == np.float32 or mask_gray.dtype == np.float64:
mask_gray = (mask_gray * 255).astype(np.uint8)
raw_gray = cast(
ImageArray,
cv2.resize(
mask_gray,
(DISPLAY_WIDTH, DISPLAY_HEIGHT),
interpolation=cv2.INTER_NEAREST,
),
)
# Normalized view preparation (without indicator)
if silhouette is None:
norm_gray = np.zeros((DISPLAY_HEIGHT, DISPLAY_WIDTH), dtype=np.uint8)
else:
upscaled = self._upscale_silhouette(silhouette)
norm_gray = upscaled
# Stack horizontally
combined = np.hstack([raw_gray, norm_gray])
# Convert back to BGR
combined_bgr = cast(ImageArray, cv2.cvtColor(combined, cv2.COLOR_GRAY2BGR))
# Add mode indicator
self._draw_mode_indicator(combined_bgr, "Both: Raw | Normalized")
return combined_bgr
def _draw_mode_indicator(
self,
image: ImageArray,
label: str,
) -> None:
"""Draw mode indicator text on image.
Args:
image: Image to draw on (modified in place)
label: Mode label text
"""
def _draw_mode_indicator(self, image: ImageArray, label: str) -> None:
h, w = image.shape[:2]
# Mode text at bottom
mode_text = f"Mode: {MODE_LABELS[self.mask_mode]} ({self.mask_mode}) - {label}"
mode_text = label
font = cv2.FONT_HERSHEY_SIMPLEX
font_scale = 0.5
@@ -383,15 +479,22 @@ class OpenCVVisualizer:
mode_text, font, font_scale, thickness
)
# Draw background at bottom center
x_pos = (w - text_width) // 2
y_pos = h - 10
x_pos = 14
y_pos = h - 8
y_top = max(0, h - MODE_LABEL_PAD)
_ = cv2.rectangle(
image,
(x_pos - 5, y_pos - text_height - 5),
(x_pos + text_width + 5, y_pos + 5),
COLOR_BLACK,
(0, y_top),
(w, h),
COLOR_DARK_GRAY,
-1,
)
_ = cv2.rectangle(
image,
(x_pos - 6, y_pos - text_height - 6),
(x_pos + text_width + 8, y_pos + 6),
COLOR_DARK_GRAY,
-1,
)
@@ -410,9 +513,11 @@ class OpenCVVisualizer:
self,
frame: ImageArray,
bbox: BBoxXYXY | None,
bbox_mask: BBoxXYXY | None,
track_id: int,
mask_raw: ImageArray | None,
silhouette: NDArray[np.float32] | None,
segmentation_input: NDArray[np.float32] | None,
label: str | None,
confidence: float | None,
fps: float,
@@ -441,23 +546,42 @@ class OpenCVVisualizer:
cv2.imshow(MAIN_WINDOW, main_display)
# Prepare and show segmentation window
seg_display = self._prepare_segmentation_view(mask_raw, silhouette)
seg_display = self._prepare_segmentation_view(mask_raw, silhouette, bbox)
cv2.imshow(SEG_WINDOW, seg_display)
if self.show_raw_window:
self._ensure_raw_window()
raw_display = self._prepare_raw_view(mask_raw, bbox_mask)
cv2.imshow(RAW_WINDOW, raw_display)
seg_input_display = self._prepare_segmentation_input_view(segmentation_input)
cv2.imshow(WINDOW_SEG_INPUT, seg_input_display)
# Handle keyboard input
key = cv2.waitKey(1) & 0xFF
if key == ord("q"):
return False
elif key == ord("m"):
# Cycle through modes: 0 -> 1 -> 2 -> 0
self.mask_mode = (self.mask_mode + 1) % 3
logger.debug("Switched to mask mode: %s", MODE_LABELS[self.mask_mode])
elif key == ord("r"):
self.show_raw_window = not self.show_raw_window
if self.show_raw_window:
self._ensure_raw_window()
logger.debug("Raw mask window enabled")
else:
self._hide_raw_window()
logger.debug("Raw mask window disabled")
elif key == ord("d"):
self.show_raw_debug = not self.show_raw_debug
logger.debug(
"Raw mask debug overlay %s",
"enabled" if self.show_raw_debug else "disabled",
)
return True
def close(self) -> None:
"""Close all OpenCV windows and cleanup."""
if self._windows_created:
self._hide_raw_window()
cv2.destroyAllWindows()
self._windows_created = False
self._raw_window_created = False
+9
View File
@@ -216,6 +216,15 @@ class SilhouetteWindow:
raise ValueError("Window is empty")
return int(self._frame_indices[0])
@property
def buffered_silhouettes(self) -> Float[ndarray, "n 64 44"]:
if not self._buffer:
return np.empty((0, SIL_HEIGHT, SIL_WIDTH), dtype=np.float32)
return cast(
Float[ndarray, "n 64 44"],
np.stack(list(self._buffer), axis=0).astype(np.float32, copy=True),
)
def _to_numpy(obj: _ArrayLike) -> ndarray:
"""Safely convert array-like object to numpy array.
View File
+394
View File
@@ -0,0 +1,394 @@
#!/usr/bin/env python3
"""
Export all positive labeled batches from Scoliosis1K dataset as time windows.
Creates grid visualizations similar to visualizer._prepare_segmentation_input_view()
for all positive class samples, arranged in sliding time windows.
Optimized UI with:
- Subject ID and batch info footer
- Dual frame counts (window-relative and sequence-relative)
- Clean layout with proper spacing
"""
from __future__ import annotations
import json
import pickle
from pathlib import Path
from typing import Final
import cv2
import numpy as np
from numpy.typing import NDArray
# Constants matching visualizer.py
DISPLAY_HEIGHT: Final = 256
DISPLAY_WIDTH: Final = 176
SIL_HEIGHT: Final = 64
SIL_WIDTH: Final = 44
# Footer settings
FOOTER_HEIGHT: Final = 80 # Height for metadata footer
FOOTER_BG_COLOR: Final = (40, 40, 40) # Dark gray background
TEXT_COLOR: Final = (255, 255, 255) # White text
ACCENT_COLOR: Final = (0, 255, 255) # Cyan for emphasis
FONT: Final = cv2.FONT_HERSHEY_SIMPLEX
FONT_SCALE: Final = 0.6
FONT_THICKNESS: Final = 2
def upscale_silhouette(
silhouette: NDArray[np.float32] | NDArray[np.uint8],
) -> NDArray[np.uint8]:
"""Upscale silhouette to display size."""
if silhouette.dtype == np.float32 or silhouette.dtype == np.float64:
sil_u8 = (silhouette * 255).astype(np.uint8)
else:
sil_u8 = silhouette.astype(np.uint8)
upscaled = cv2.resize(
sil_u8, (DISPLAY_WIDTH, DISPLAY_HEIGHT), interpolation=cv2.INTER_NEAREST
)
return upscaled
def create_optimized_visualization(
silhouettes: NDArray[np.float32],
subject_id: str,
view_name: str,
window_idx: int,
start_frame: int,
end_frame: int,
n_frames_total: int,
tile_height: int = DISPLAY_HEIGHT,
tile_width: int = DISPLAY_WIDTH,
) -> NDArray[np.uint8]:
"""
Create optimized visualization with grid and metadata footer.
Args:
silhouettes: Array of shape (n_frames, 64, 44) float32
subject_id: Subject identifier
view_name: View identifier (e.g., "000_180")
window_idx: Window index within sequence
start_frame: Starting frame index in sequence
end_frame: Ending frame index in sequence
n_frames_total: Total frames in the sequence
tile_height: Height of each tile in the grid
tile_width: Width of each tile in the grid
Returns:
Combined image with grid visualization and metadata footer
"""
n_frames = int(silhouettes.shape[0])
tiles_per_row = int(np.ceil(np.sqrt(n_frames)))
rows = int(np.ceil(n_frames / tiles_per_row))
# Create grid
grid = np.zeros((rows * tile_height, tiles_per_row * tile_width), dtype=np.uint8)
# Place each silhouette in the grid
for idx in range(n_frames):
sil = silhouettes[idx]
tile = upscale_silhouette(sil)
r = idx // tiles_per_row
c = idx % tiles_per_row
y0, y1 = r * tile_height, (r + 1) * tile_height
x0, x1 = c * tile_width, (c + 1) * tile_width
grid[y0:y1, x0:x1] = tile
# Convert to BGR
grid_bgr = cv2.cvtColor(grid, cv2.COLOR_GRAY2BGR)
# Add frame indices as text (both window-relative and sequence-relative)
for idx in range(n_frames):
r = idx // tiles_per_row
c = idx % tiles_per_row
y0 = r * tile_height
x0 = c * tile_width
# Window frame count (top-left)
cv2.putText(
grid_bgr,
f"{idx}", # Window-relative frame number
(x0 + 8, y0 + 22),
FONT,
FONT_SCALE,
ACCENT_COLOR,
FONT_THICKNESS,
cv2.LINE_AA,
)
# Sequence frame count (bottom-left of tile)
seq_frame = start_frame + idx
cv2.putText(
grid_bgr,
f"#{seq_frame}", # Sequence-relative frame number
(x0 + 8, y0 + tile_height - 10),
FONT,
0.45, # Slightly smaller font
(180, 180, 180), # Light gray
1,
cv2.LINE_AA,
)
# Create footer with metadata
grid_width = grid_bgr.shape[1]
footer = np.full((FOOTER_HEIGHT, grid_width, 3), FOOTER_BG_COLOR, dtype=np.uint8)
# Line 1: Subject ID and view
line1 = f"Subject: {subject_id} | View: {view_name}"
cv2.putText(
footer,
line1,
(15, 25),
FONT,
0.7,
TEXT_COLOR,
FONT_THICKNESS,
cv2.LINE_AA,
)
# Line 2: Window batch frame range
line2 = f"Window {window_idx}: frames [{start_frame:03d} - {end_frame - 1:03d}] ({n_frames} frames)"
cv2.putText(
footer,
line2,
(15, 50),
FONT,
0.7,
ACCENT_COLOR,
FONT_THICKNESS,
cv2.LINE_AA,
)
# Line 3: Progress within sequence
progress_pct = (end_frame / n_frames_total) * 100
line3 = f"Sequence: {n_frames_total} frames total | Progress: {progress_pct:.1f}%"
cv2.putText(
footer,
line3,
(15, 72),
FONT,
0.6,
(200, 200, 200),
1,
cv2.LINE_AA,
)
# Combine grid and footer
combined = np.vstack([grid_bgr, footer])
return combined
def load_pkl_sequence(pkl_path: Path) -> NDArray[np.float32]:
"""Load a .pkl file containing silhouette sequence."""
with open(pkl_path, "rb") as f:
data = pickle.load(f)
# Handle different possible structures
if isinstance(data, np.ndarray):
return data.astype(np.float32)
elif isinstance(data, list):
# List of frames
return np.stack([np.array(frame) for frame in data]).astype(np.float32)
else:
raise ValueError(f"Unexpected data type in {pkl_path}: {type(data)}")
def create_windows(
sequence: NDArray[np.float32],
window_size: int = 30,
stride: int = 30,
) -> list[NDArray[np.float32]]:
"""
Split a sequence into sliding windows.
Args:
sequence: Array of shape (N, 64, 44)
window_size: Number of frames per window
stride: Stride between consecutive windows
Returns:
List of window arrays, each of shape (window_size, 64, 44)
"""
n_frames = sequence.shape[0]
windows = []
for start_idx in range(0, n_frames - window_size + 1, stride):
end_idx = start_idx + window_size
window = sequence[start_idx:end_idx]
windows.append(window)
return windows
def export_positive_batches(
dataset_root: Path,
output_dir: Path,
window_size: int = 30,
stride: int = 30,
max_sequences: int | None = None,
) -> None:
"""
Export all positive labeled batches from Scoliosis1K dataset as time windows.
Args:
dataset_root: Path to Scoliosis1K-sil-pkl directory
output_dir: Output directory for visualizations
window_size: Number of frames per window (default 30)
stride: Stride between consecutive windows (default 30 = non-overlapping)
max_sequences: Maximum number of sequences to process (None = all)
"""
output_dir.mkdir(parents=True, exist_ok=True)
# Find all positive samples
positive_samples: list[
tuple[Path, str, str, str]
] = [] # (pkl_path, subject_id, view_name, pkl_name)
for subject_dir in sorted(dataset_root.iterdir()):
if not subject_dir.is_dir():
continue
subject_id = subject_dir.name
# Check for positive class directory (lowercase)
positive_dir = subject_dir / "positive"
if not positive_dir.exists():
continue
# Iterate through views
for view_dir in sorted(positive_dir.iterdir()):
if not view_dir.is_dir():
continue
view_name = view_dir.name
# Find .pkl files
for pkl_file in sorted(view_dir.glob("*.pkl")):
positive_samples.append(
(pkl_file, subject_id, view_name, pkl_file.stem)
)
print(f"Found {len(positive_samples)} positive labeled sequences")
if max_sequences:
positive_samples = positive_samples[:max_sequences]
print(f"Processing first {max_sequences} sequences")
total_windows = 0
# Export each sequence's windows
for seq_idx, (pkl_path, subject_id, view_name, pkl_name) in enumerate(
positive_samples, 1
):
print(
f"[{seq_idx}/{len(positive_samples)}] Processing {subject_id}/{view_name}/{pkl_name}..."
)
# Load sequence
try:
sequence = load_pkl_sequence(pkl_path)
except Exception as e:
print(f" Error loading {pkl_path}: {e}")
continue
# Ensure correct shape (N, 64, 44)
if len(sequence.shape) == 2:
# Single frame
sequence = sequence[np.newaxis, ...]
elif len(sequence.shape) == 3:
# (N, H, W) - expected
pass
else:
print(f" Unexpected shape {sequence.shape}, skipping")
continue
n_frames = sequence.shape[0]
print(f" Sequence has {n_frames} frames")
# Skip if sequence is shorter than window size
if n_frames < window_size:
print(f" Skipping: sequence too short (< {window_size} frames)")
continue
# Normalize if needed
if sequence.max() > 1.0:
sequence = sequence / 255.0
# Create windows
windows = create_windows(sequence, window_size=window_size, stride=stride)
print(f" Created {len(windows)} windows (size={window_size}, stride={stride})")
# Export each window
for window_idx, window in enumerate(windows):
start_frame = window_idx * stride
end_frame = start_frame + window_size
# Create visualization for this window with full metadata
vis_image = create_optimized_visualization(
silhouettes=window,
subject_id=subject_id,
view_name=view_name,
window_idx=window_idx,
start_frame=start_frame,
end_frame=end_frame,
n_frames_total=n_frames,
)
# Save with descriptive filename including window index
output_filename = (
f"{subject_id}_{view_name}_{pkl_name}_win{window_idx:03d}.png"
)
output_path = output_dir / output_filename
cv2.imwrite(str(output_path), vis_image)
# Save metadata for this window
meta = {
"subject_id": subject_id,
"view": view_name,
"pkl_name": pkl_name,
"window_index": window_idx,
"window_size": window_size,
"stride": stride,
"start_frame": start_frame,
"end_frame": end_frame,
"sequence_shape": sequence.shape,
"n_frames_total": n_frames,
"source_path": str(pkl_path),
}
meta_filename = (
f"{subject_id}_{view_name}_{pkl_name}_win{window_idx:03d}.json"
)
meta_path = output_dir / meta_filename
with open(meta_path, "w") as f:
json.dump(meta, f, indent=2)
total_windows += 1
print(f" Exported {len(windows)} windows")
print(f"\nExport complete! Saved {total_windows} windows to {output_dir}")
def main() -> None:
"""Main entry point."""
# Paths
dataset_root = Path("/mnt/public/data/Scoliosis1K/Scoliosis1K-sil-pkl")
output_dir = Path("/home/crosstyan/Code/OpenGait/output/positive_batches")
if not dataset_root.exists():
print(f"Error: Dataset not found at {dataset_root}")
return
# Export all positive batches with windowing
export_positive_batches(
dataset_root,
output_dir,
window_size=30, # 30 frames per window
stride=30, # Non-overlapping windows
)
if __name__ == "__main__":
main()
+64 -8
View File
@@ -7,7 +7,7 @@ from pathlib import Path
import subprocess
import sys
import time
from typing import Final, cast
from typing import Final, Literal, cast
from unittest import mock
import numpy as np
@@ -693,9 +693,11 @@ class MockVisualizer:
self,
frame: NDArray[np.uint8],
bbox: tuple[int, int, int, int] | None,
bbox_mask: tuple[int, int, int, int] | None,
track_id: int,
mask_raw: NDArray[np.uint8] | None,
silhouette: NDArray[np.float32] | None,
segmentation_input: NDArray[np.float32] | None,
label: str | None,
confidence: float | None,
fps: float,
@@ -704,9 +706,11 @@ class MockVisualizer:
{
"frame": frame,
"bbox": bbox,
"bbox_mask": bbox_mask,
"track_id": track_id,
"mask_raw": mask_raw,
"silhouette": silhouette,
"segmentation_input": segmentation_input,
"label": label,
"confidence": confidence,
"fps": fps,
@@ -761,9 +765,8 @@ def test_pipeline_visualizer_updates_on_no_detection() -> None:
visualize=True,
)
# Replace the visualizer with our mock
mock_viz = MockVisualizer()
pipeline._visualizer = mock_viz # type: ignore[assignment]
setattr(pipeline, "_visualizer", mock_viz)
# Run pipeline
_ = pipeline.run()
@@ -779,13 +782,14 @@ def test_pipeline_visualizer_updates_on_no_detection() -> None:
for call in mock_viz.update_calls:
assert call["track_id"] == 0 # Default track_id when no detection
assert call["bbox"] is None # No bbox when no detection
assert call["bbox_mask"] is None
assert call["mask_raw"] is None # No mask when no detection
assert call["silhouette"] is None # No silhouette when no detection
assert call["segmentation_input"] is None
assert call["label"] is None # No label when no detection
assert call["confidence"] is None # No confidence when no detection
def test_pipeline_visualizer_uses_cached_detection_on_no_detection() -> None:
"""Test that visualizer reuses last valid detection when current frame has no detection.
@@ -835,7 +839,12 @@ def test_pipeline_visualizer_uses_cached_detection_on_no_detection() -> None:
dummy_mask = np.random.randint(0, 256, (480, 640), dtype=np.uint8)
dummy_bbox_mask = (100, 100, 200, 300)
dummy_bbox_frame = (100, 100, 200, 300)
mock_select_person.return_value = (dummy_mask, dummy_bbox_mask, dummy_bbox_frame, 1)
mock_select_person.return_value = (
dummy_mask,
dummy_bbox_mask,
dummy_bbox_frame,
1,
)
# Setup mock mask_to_silhouette to return valid silhouette
dummy_silhouette = np.random.rand(64, 44).astype(np.float32)
@@ -856,9 +865,8 @@ def test_pipeline_visualizer_uses_cached_detection_on_no_detection() -> None:
visualize=True,
)
# Replace the visualizer with our mock
mock_viz = MockVisualizer()
pipeline._visualizer = mock_viz # type: ignore[assignment]
setattr(pipeline, "_visualizer", mock_viz)
# Run pipeline
_ = pipeline.run()
@@ -886,9 +894,57 @@ def test_pipeline_visualizer_uses_cached_detection_on_no_detection() -> None:
"not None/blank"
)
# The cached masks should be copies (different objects) to prevent mutation issues
segmentation_inputs = [
call["segmentation_input"] for call in mock_viz.update_calls
]
bbox_mask_calls = [call["bbox_mask"] for call in mock_viz.update_calls]
assert segmentation_inputs[0] is not None
assert segmentation_inputs[1] is not None
assert segmentation_inputs[2] is not None
assert segmentation_inputs[3] is not None
assert bbox_mask_calls[0] == dummy_bbox_mask
assert bbox_mask_calls[1] == dummy_bbox_mask
assert bbox_mask_calls[2] == dummy_bbox_mask
assert bbox_mask_calls[3] == dummy_bbox_mask
if mask_raw_calls[1] is not None and mask_raw_calls[2] is not None:
assert mask_raw_calls[1] is not mask_raw_calls[2], (
"Cached mask should be a copy, not the same object reference"
)
def test_frame_pacer_emission_count_24_to_15() -> None:
from opengait.demo.pipeline import _FramePacer
pacer = _FramePacer(15.0)
interval_ns = int(1_000_000_000 / 24)
emitted = sum(pacer.should_emit(i * interval_ns) for i in range(100))
assert 60 <= emitted <= 65
def test_frame_pacer_requires_positive_target_fps() -> None:
from opengait.demo.pipeline import _FramePacer
with pytest.raises(ValueError, match="target_fps must be positive"):
_FramePacer(0.0)
@pytest.mark.parametrize(
("window", "stride", "mode", "expected"),
[
(30, 30, "manual", 30),
(30, 7, "manual", 7),
(30, 30, "sliding", 1),
(30, 1, "chunked", 30),
(15, 3, "chunked", 15),
],
)
def test_resolve_stride_modes(
window: int,
stride: int,
mode: Literal["manual", "sliding", "chunked"],
expected: int,
) -> None:
from opengait.demo.pipeline import resolve_stride
assert resolve_stride(window, stride, mode) == expected
+171
View File
@@ -0,0 +1,171 @@
from __future__ import annotations
from pathlib import Path
from typing import cast
from unittest import mock
import numpy as np
import pytest
from opengait.demo.input import create_source
from opengait.demo.visualizer import (
DISPLAY_HEIGHT,
DISPLAY_WIDTH,
ImageArray,
OpenCVVisualizer,
)
from opengait.demo.window import select_person
REPO_ROOT = Path(__file__).resolve().parents[2]
SAMPLE_VIDEO_PATH = REPO_ROOT / "assets" / "sample.mp4"
YOLO_MODEL_PATH = REPO_ROOT / "ckpt" / "yolo11n-seg.pt"
def test_prepare_raw_view_float_mask_has_visible_signal() -> None:
viz = OpenCVVisualizer()
mask_float = np.zeros((64, 64), dtype=np.float32)
mask_float[16:48, 16:48] = 1.0
rendered = viz._prepare_raw_view(cast(ImageArray, mask_float))
assert rendered.dtype == np.uint8
assert rendered.shape == (256, 176, 3)
mask_zero = np.zeros((64, 64), dtype=np.float32)
rendered_zero = viz._prepare_raw_view(cast(ImageArray, mask_zero))
roi = slice(0, DISPLAY_HEIGHT - 40)
diff = np.abs(rendered[roi].astype(np.int16) - rendered_zero[roi].astype(np.int16))
assert int(np.count_nonzero(diff)) > 0
def test_prepare_raw_view_handles_values_slightly_above_one() -> None:
viz = OpenCVVisualizer()
mask = np.zeros((64, 64), dtype=np.float32)
mask[20:40, 20:40] = 1.0001
rendered = viz._prepare_raw_view(cast(ImageArray, mask))
roi = rendered[: DISPLAY_HEIGHT - 40, :, 0]
assert int(np.count_nonzero(roi)) > 0
def test_segmentation_view_is_normalized_only_shape() -> None:
viz = OpenCVVisualizer()
mask = np.zeros((480, 640), dtype=np.uint8)
sil = np.random.rand(64, 44).astype(np.float32)
seg = viz._prepare_segmentation_view(cast(ImageArray, mask), sil, (0, 0, 100, 100))
assert seg.shape == (DISPLAY_HEIGHT, DISPLAY_WIDTH, 3)
def test_update_toggles_raw_window_with_r_key() -> None:
viz = OpenCVVisualizer()
frame = np.zeros((240, 320, 3), dtype=np.uint8)
mask = np.zeros((240, 320), dtype=np.uint8)
mask[20:100, 30:120] = 255
sil = np.random.rand(64, 44).astype(np.float32)
seg_input = np.random.rand(4, 64, 44).astype(np.float32)
with (
mock.patch("cv2.namedWindow") as named_window,
mock.patch("cv2.imshow"),
mock.patch("cv2.destroyWindow") as destroy_window,
mock.patch("cv2.waitKey", side_effect=[ord("r"), ord("r"), ord("q")]),
):
assert viz.update(
frame,
(10, 10, 120, 150),
(10, 10, 120, 150),
1,
cast(ImageArray, mask),
sil,
seg_input,
None,
None,
15.0,
)
assert viz.show_raw_window is True
assert viz._raw_window_created is True
assert viz.update(
frame,
(10, 10, 120, 150),
(10, 10, 120, 150),
1,
cast(ImageArray, mask),
sil,
seg_input,
None,
None,
15.0,
)
assert viz.show_raw_window is False
assert viz._raw_window_created is False
assert destroy_window.called
assert (
viz.update(
frame,
(10, 10, 120, 150),
(10, 10, 120, 150),
1,
cast(ImageArray, mask),
sil,
seg_input,
None,
None,
15.0,
)
is False
)
assert named_window.called
def test_sample_video_raw_mask_shape_range_and_render_signal() -> None:
if not SAMPLE_VIDEO_PATH.is_file():
pytest.skip(f"Missing sample video: {SAMPLE_VIDEO_PATH}")
if not YOLO_MODEL_PATH.is_file():
pytest.skip(f"Missing YOLO model file: {YOLO_MODEL_PATH}")
ultralytics = pytest.importorskip("ultralytics")
yolo_cls = getattr(ultralytics, "YOLO")
viz = OpenCVVisualizer()
detector = yolo_cls(str(YOLO_MODEL_PATH))
masks_seen = 0
rendered_nonzero: list[int] = []
for frame, _meta in create_source(str(SAMPLE_VIDEO_PATH), max_frames=30):
detections = detector.track(
frame,
persist=True,
verbose=False,
classes=[0],
device="cpu",
)
if not isinstance(detections, list) or not detections:
continue
selected = select_person(detections[0])
if selected is None:
continue
mask_raw, _, _, _ = selected
masks_seen += 1
arr = np.asarray(mask_raw)
assert arr.ndim == 2
assert arr.shape[0] > 0 and arr.shape[1] > 0
assert np.issubdtype(arr.dtype, np.number)
assert float(arr.min()) >= 0.0
assert float(arr.max()) <= 255.0
assert int(np.count_nonzero(arr)) > 0
rendered = viz._prepare_raw_view(arr)
roi = rendered[: DISPLAY_HEIGHT - 40, :, 0]
rendered_nonzero.append(int(np.count_nonzero(roi)))
assert masks_seen > 0
assert min(rendered_nonzero) > 0