Use buffered frame indices for emitted window bounds to stay accurate across detection gaps, and align select_person tests with the 4-field return contract introduced for frame-space bbox support.
Align bbox coordinate handling across primary and fallback paths, normalize Both-mode raw mask rendering, and tighten demo result typing to reduce runtime/display inconsistencies.
Apply Oracle-guided cleanup to make the demo pipeline contract explicit and remove defensive runtime indirection while preserving existing visualization behavior.
Integrate an opt-in OpenCV visualizer into the demo runtime so operators can monitor tracking, segmentation, and inference confidence in real time without changing the default non-visual execution path.
Add preprocess-only silhouette export and configurable result exporters so demo runs can be persisted for offline analysis and reproducible evaluation. Include optional parquet support and CLI visualization dumps while updating tests and tracking notes for the verified pipeline/debug workflow.
Add the full demo runtime stack for single-person scoliosis inference, including input adapters, silhouette preprocessing, temporal windowing, ScoNet wrapper, result publishing, and click-based CLI orchestration. This commit captures the executable pipeline behavior independently from tests and planning artifacts for clearer review and rollback.
- ScoNet:
- Previously, `label_ids` remained a NumPy array, which could cause dtype/device mismatches when used with PyTorch tensors on GPU.
- Convert `label_ids` to `torch.from_numpy(...).cuda().long()` to ensure correct tensor type (Long) and device (CUDA), aligning with loss functions that expect class indices on the same device.
crosstyan
Zzier
bugjudger
Dongyang Jin
chuanfushen
Noah
darkliang
Chao Fan
Jingzhe Ma
jdyjjj