Supplementary material for submission 18051. Highlighted text marks items that are still being finalised.
Rebuttal summary
General response
We thank all reviewers. All four recommend acceptance (weak accept). Most questions ask for analyses of results we already
have, plus a few new runs. New in this response:
Retrieval hit rate and training distribution for 100 / 200 / 400 s (btN3). Longer spans add *available* first visits details
(83% -> 91% -> 96% of scored chunks have one in range); given one is available, top-6 retrieves it about 90% of the time at every span.
Camera-following on all five rollout seeds, with CaR, and the return comparison stratified by realized yaw error (btN3). details
Among returns the camera actually makes (error <= 10 deg), CaR is within seed noise of Memorizon 100 s / 400 s and 0.06 Revisit below 200 s; the other open models stay far below.
We report this plainly and soften the cross-system claim accordingly.
Cost (m2mq, MjTC). Pose scoring takes 32 ms per chunk at 400 s of history on one CPU thread, about 1% of the 3.2 s it takes to generate a chunk; chunk latency is flat from 10 s to 400 s of history (3.18 -> 3.19 s) and peak allocated GPU memory is identical to a sliding window (22.6 GB). details
Bank-content ablation on the 200 s and 400 s models (no bank / random / foreign bank: Gain falls to about 0) and independent training seeds for 200 s and 400 s [PENDING: G2], and bank noise for the 400 s quality drop [PENDING: G1]. details
Zero-shot on real capture (KgbX, m2mq): 20 DL3DV-10K scenes from a single real first frame and the scene's own camera path; path following holds (correlation 0.79 / 0.73), pixel fidelity to the real frames is not claimed. Videos on the supplementary page. details
Occlusion of retrieved frames (MjTC), with monocular depth: 24-49% of the in-frustum content of retrieved frames is occluded versus 38-65% for random frames passing the same frustum test. details
Wording fixes: "bounded by k x K" in the abstract; the 400 s explanation is now a hypothesis.