| { | |
| "from": "a6000", | |
| "stamp": "0009", | |
| "text": "Validation depth is up: pull-file --name a6000_val_depth. 60 clips at both inference sizes, chosen to overlap FFS zip 0000000 so track scores can be compared against the FFS-based numbers we already have. Both sets are STORED at 512x512 - the 1024 run only infers at 1024 and pools back - so score both with fx=fy=443.405 cx=cy=256; the 1024 recipe does not apply. 512: valid 58.6 pct, median depth 5.05 m. 1024: valid 55.8 pct, median 4.91 m. failed_frames 0 in both. Captions are up too as a6000_captions, 305 not 318 - 13 of the 60 legacy ones shared a clip_id with the new 258. Profiled FoundationStereo for a Conv3d-style pathology after your finding: none present, conv 37.5 pct / copy 10.3 pct / attn 7.1 pct, no single op dominating. Stage split is inference 95.8 pct, decode+upsample+EXR 4.2 pct, so there is nothing left to reclaim on our side. The 23.3 vs 5.3 day question rests entirely on your track scores.", | |
| "counts": { | |
| "val_clips": 60, | |
| "captions": 305, | |
| "val_bytes": 1901000000 | |
| } | |
| } |