ChangChrisLiu commited on
Commit
1bfcccd
·
verified ·
1 Parent(s): 3b76a5f

Add Episode 3 (196 frames, goal=graphic_card) + v3.1 two-flag encoding notes

Browse files
Files changed (1) hide show
  1. README.md +16 -3
README.md CHANGED
@@ -251,10 +251,23 @@ Two episodes are fully labeled and validated — you can use them to test the lo
251
  | Robot visibility | Visible in 448 / 473 frames |
252
  | Goal component | `cpu_fan` |
253
 
254
- **Note:** Episode 2 uses the v3.1 labeling pipeline which includes:
 
 
 
 
 
 
 
 
 
 
 
 
255
  - **Amodal masks** — stored masks represent the full component shape, unchanged when robot occludes
256
- - **Robot-aware depth** — at save time, robot mask is subtracted before computing 3D depth features, so `centroid`/`point_cloud` sample only non-occluded pixels. When fully occluded, `depth_valid=0` (but `visibility=1` — component still exists).
257
- - **SAM2 goal tracking** the goal component is SAM2-tracked across frames instead of naive mask copy.
 
258
 
259
  ### Download and Test (3 steps)
260
 
 
251
  | Robot visibility | Visible in 448 / 473 frames |
252
  | Goal component | `cpu_fan` |
253
 
254
+ ### Episode 3: `session_0410_125013/episode_01`
255
+
256
+ Continuation episode — cpu_fan was already removed in Episode 2, so this episode starts with all remaining components visible.
257
+
258
+ | Stat | Value |
259
+ |------|-------|
260
+ | Total frames in episode | 196 |
261
+ | Labeled frames | **196** (range 0–195, **0 gaps, 100% coverage**) |
262
+ | Product components | 11 (cpu_bracket, cpu, graphic_card, motherboard, connector_1..4, ram_1, ram_clip_1..2) |
263
+ | Physical constraints (edges) | 11 |
264
+ | Goal component | `graphic_card` |
265
+
266
+ **Note on v3.1 data semantics** (applies to all labeled frames):
267
  - **Amodal masks** — stored masks represent the full component shape, unchanged when robot occludes
268
+ - **Robot-aware depth and embeddings** — at save time, `mask & ~robot_mask` is used for `compute_depth_info` and embedding re-extraction, so `centroid`, `point_cloud`, and 256-D features sample only non-occluded pixels. When fully occluded, `depth_valid=0` and embedding is zeroed (but `visibility=1` — component still in scene, just unmeasurable this frame).
269
+ - **Two-flag semantic encoding**: `(visibility, depth_valid)` = `(1,1)` visible & measurable; `(1,0)` fully robot-occluded; `(0,0)` not in scene (hidden or removed — disambiguated via incoming edge `is_locked`).
270
+ - **SAM2 tracking on copy** — goal component and robot get SAM2 centroid+bbox+mask-input prediction across frames; other components simple-copy.
271
 
272
  ### Download and Test (3 steps)
273