7. Occlusion of Retrieved Frames

Motivation

Reviewer MjTC: "How does the model behave in multi-room scenes when frustums overlap through walls?"

Results

Frustum overlap ignores occlusion (Appendix A.4). Measured: retrieved frames are less occluded than random frames that pass the same test.

Table 1: Occluded share of in-frustum content, retrieved vs random frames (monocular depth).

depth tolerance	retrieved, occluded share	random, occluded share
τ = 0.25	24%	38%
τ = 0.15	37%	53%
τ = 0.10	49%	65%

399 chunks sampled as in training; occluded share is the part of in-frustum content that is hidden in the candidate.
For 79–94% of chunks with a frustum hit, at least one retrieved frame has visible overlap ≥ 0.3.
Limits: depth is monocular (Depth Anything 3), not ground truth; τ is our choice. Indirect evidence only. [VERIFY wording]

Examples

First 8 sampled chunks, no selection. Top row: query frame, retrieved frame, query overlay. Bottom row: random control. Green = same surface visible in the candidate, red = occluded by another depth.

Chunk 1: retrieved occluded 42%, random occluded 42%

Chunk 3: retrieved occluded 51%, random occluded 34%

Chunk 4: retrieved occluded 47%, random occluded 63%

Show the other 5

Chunk 5: retrieved occluded 8%, random occluded 42%

Chunk 6: retrieved occluded 14%, random occluded 24%

Chunk 7: retrieved occluded 45%, random occluded 8%

Chunk 8: retrieved occluded 12%, random occluded 30%

Chunk 9: retrieved occluded 27%, random occluded 1%

Over these 8 chunks the two are equal (30.5% vs 30.4%; retrieved is lower in 4 of 8). The gap in the table comes from all 399 chunks, not from these examples.

Conclusion

Retrieved frames are less occluded than random frames that pass the same frustum test: 24–49% vs 38–65% of in-frustum content, depending on tolerance.
Most chunks have a usable retrieved frame: 79–94% of chunks with a frustum hit have one with visible overlap ≥ 0.3.
Indirect evidence only: depth is monocular, and τ is our choice.