Promote runtime deps and archive scoliosis analysis

2026-03-10 15:02:46 +08:00
parent 560f8e7aae
commit 2647398307
3 changed files with 21 additions and 20 deletions
@@ -0,0 +1,74 @@
+# Scoliosis1K Dataset Analysis (1:1:8, shared-align skeleton maps)
+
+## Split
+
+Train counts:
+- negative: 596
+- neutral: 74
+- positive: 74
+
+Test counts:
+- negative: 204
+- neutral: 126
+- positive: 419
+
+## Raw pose quality
+
+Mean keypoint confidence by split/class:
+- train negative: 0.9016
+- train neutral: 0.9023
+- train positive: 0.8987
+- test negative: 0.9009
+- test neutral: 0.9020
+- test positive: 0.8999
+
+Mean valid-joint ratio (`conf > 0.05`) by split/class:
+- train negative: 1.0000
+- train neutral: 1.0000
+- train positive: 1.0000
+- test negative: 1.0000
+- test neutral: 1.0000
+- test positive: 1.0000
+
+## PAV signal
+
+Mean normalized PAV value by label:
+- negative: 0.3068
+- neutral: 0.3546
+- positive: 0.3635
+
+Train-on-train / test-on-test linear softmax probe over sequence-level PAV:
+- accuracy: 50.87%
+- macro precision: 50.50%
+- macro recall: 48.19%
+- macro F1: 39.88%
+
+## Shared-align heatmap geometry
+
+Combined support bbox stats over all sequences:
+- width mean / p95: 32.13 / 33.57
+- height mean / p95: 61.52 / 61.61
+- active fraction mean / p95: 0.3634 / 0.3738
+
+Per-sequence temporal jitter (std over frames):
+- center-x std mean / p95: 0.864 / 1.243
+- center-y std mean / p95: 0.516 / 0.704
+- width std mean / p95: 2.152 / 2.804
+- height std mean / p95: 0.507 / 0.545
+
+Residual limb-vs-joint bbox-center mismatch after shared alignment:
+- dx mean / p95: 0.195 / 0.229
+- dy mean / p95: 0.251 / 0.357
+
+Estimated intensity mass in the columns removed by `BaseSilCuttingTransform`:
+- mean clipped-mass ratio: 0.0000
+- p95 clipped-mass ratio: 0.0000
+
+## Reading
+
+- The raw pose data does not look broken. Confidence and valid-joint ratios are high and similar across classes.
+- The sequence-level PAV still carries useful label signal, so the dataset is not devoid of scoliosis information.
+- Shared alignment removed the old limb-vs-joint registration bug; residual channel-center mismatch is now small.
+- The remaining suspicious area is the visual branch: the skeleton map still has frame-to-frame bbox jitter, and the support bbox is almost full-height (`~61.5 / 64`) and fairly dense (`~36%` active pixels), which may be washing out subtle asymmetry cues.
+- `BaseSilCuttingTransform` does not appear to be the main failure source for this shared-align export; the measured mass in the removed side margins is near zero.
+- The dataset itself looks usable; the bigger issue still appears to be how the current skeleton-map preprocessing/runtime path presents that data to ScoNet.