Motivation
Reviewer btN3: "For the existing 100-, 200-, and 400-second configurations, can you report how often scored chunks actually retrieve a relevant first visit, rather than only whether that visit lies within the maximum span?"
Results
Table 1: Retrieval hit rate by maximum training span. Relevant means the latent sees at least half of the chunk's view frustum.
| max span | first visit exists | top-6 hits one | hit, given exists | median age | older than 30 s | spans ≤ 30 s |
|---|---|---|---|---|---|---|
| 100 s | 82.7% | 76.4% | 92.5% | 16.8 s | 27% | 22.6% |
| 200 s | 91.0% | 82.6% | 90.7% | 24.5 s | 44% | 10.2% |
| 400 s | 95.5% | 86.3% | 90.3% | 37.0 s | 55% | 5.0% |
Setup
- Training sampler replayed on 380 segments, 4,000 spans per setting, ten scored chunks each.
- A latent is relevant if it sees at least half of the chunk's view frustum (0.3 and 0.7 give the same picture). Pre-window history only.
Conclusion
- Longer spans give the model more to retrieve. A relevant first visit exists for 83% → 96% of scored chunks, and retrieved frames get older (median 16.8 s → 37.0 s).
- Retrieval itself does not improve. Given a relevant visit exists, top-6 finds it about 90% of the time at every span (92.5%, 90.7%, 90.3%).
- The 400 s drop is unexplained. It may come from coverage, or from the span mix (spans ≤ 30 s fall from 10.2% to 5.0%); we cannot separate the two. Sec. 4.3 now says 200 s is best *in our tested setup*.