occlusion recovery demo (Approach A single-object re-seed) + honest caveats

occlusion_recovery_demo/README.md

@@ -0,0 +1,22 @@
+# Occlusion recovery demo (Approach A, single-object re-seed)
+
+`occlusion_recovery_demo.mp4` — pink + blue bowls tracked; the **pink** bowl is
+occluded (grey box) so its SAM2 mask vanishes, then reappears. The single-object
+recovery fires: re-detect (OWLv2) → `add_new_points_or_box` at the current frame
+for the pink bowl only → mask re-acquired. The **blue** bowl is never touched
+(that is the point of single-object recovery vs the old full reinit).
+
+## Honest caveats
+- This is a SYNTHETIC occlusion. SAM2 is very robust to clean synthetic occlusion:
+  when the *pixel-identical* frame returns it re-locks by appearance on its own —
+  even after 12 black frames — so a "without-recovery fails" contrast is NOT
+  reproducible synthetically. The video shows the recovery MECHANISM firing, not a
+  case where SAM2 would have failed.
+- The mechanism itself is verified numerically by a controlled black-out test:
+  bowl area 3490 → black frames drive it to 0 → real frame back → recovery
+  re-acquires to 3490 (and the reseed produces a valid mask).
+- The real failure the recovery targets (mask tracking lost during occlusion) is a
+  REAL-robot phenomenon: the object's appearance/position drifts during occlusion
+  (grasped + rotated, gripper alongside, lighting), which clean sim frames can't
+  reproduce. Definitive validation is on the real robot with
+  `mask_backend: groundingdino` + `detector: owlv2` (watch for `[occlusion-recovery/A]`).

occlusion_recovery_demo/occlusion_recovery_demo.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bb2b087ec40958574bb0f46d8f759affe201e9b8387343cdea6cf01a2ace4cc
+size 44596