| # frames/ — frame cache for steps 2 and 3 (companion repo GMLRVigil/BenchCheck-Pool-Frames, public) |
|
|
| Layout after unpacking: `frames/pool_frames/<video_id>/frame_<k>.jpg` (JPEG, long side 448 px) plus |
| `frames.json` (grid positions, keys, mid_k = the single frame used by step 2, w, h). One directory per |
| normalized video referenced by a step-1 item. 16 tar shards `frames_<0-f>.tar`, sharded by the first |
| hex character of video_id; `frames_manifest.json` lists sha256 and counts. `coverage.json` gives the |
| snapshot coverage: videos that were not normalized on the origin cluster at snapshot time have |
| video_id "k:..." in the items table and no frames here; steps 2/3 skip them (they are finished on the |
| origin cluster later). |
| |
