Update README with hybrid occlusion-aware pose estimation details

README.md

@@ -23,7 +23,13 @@ pretty_name: Dynamic Intelligence - Egocentric Human Motion Annotation Dataset
 
 RGB-D hand manipulation dataset captured with iPhone 13 TrueDepth sensor for humanoid robot training. Includes 6-DoF hand pose trajectories, synchronized video, and semantic motion annotations.
 
-**Data Integrity:** When hand tracking fails (out-of-frame, occlusion), states hold their last known value (frozen). Downstream models can detect this by checking for identical consecutive frames (`frame[i] == frame[i-1]`). Euler angles are unwrapped for continuous visualization without ±180° discontinuities.
+**Data Integrity & Occlusion Handling:**
+- **Visible joints:** Direct depth measurement from TrueDepth sensor (<3mm accuracy)
+- **Self-occluded joints:** Estimated via scaled MediaPipe world landmarks anchored to visible wrist position (~5-8mm accuracy)
+- **Euler angles:** Unwrapped for continuous visualization without ±180° discontinuities
+- **Confidence scores:** Each joint has a confidence value (1.0 = depth-based, 0.7 = estimated)
+
+This hybrid approach ensures no "fake" depth values are used when joints are behind the palm during grasping movements.
 
 ---