# Orbit-benchmark evaluation data: dl3dv_val33 and dl3dv_test33

`orbit_eval_data.zip` (3.39 GiB, stored uncompressed; paths relative to the repo root, so `unzip -o orbit_eval_data.zip` inside a
checkout restores them in place) - the two 33-scene slices of the DL3DV orbit (revisit) benchmark used for every number in
`09_voxel_update_f1_harm/` and `10_fusion_head_frozen_3k/`:

- `data/orbit_cache/dl3dv_test33/` (33 records + meta.json): **reported once**. The 33 DL3DV test-split videos whose cameras
  orbit a centre facing inward, at the train thresholds (>= 400 degrees around the centre, >= 60 % of views facing it); it
  contains the 8 scenes of the original benchmark (`data/orbit_cache/dl3dv_test`). Scene list: `data/orbit_cache/test_sel_loose.json`.
- `data/orbit_cache/dl3dv_val33/` (33 records + meta.json): **chooses configurations**. 33 random qualifying DL3DV train-split
  orbits (seed 20261029) never used to train the voxel update heads or the fusion head (not in `dl3dv_train`); the base trunk
  may have seen these scenes as training clips. Scene list: `data/orbit_cache/val_sel_33.json`. (On the machine that built it,
  the folder links into `data/orbit_cache/dl3dv_train_all/`; here the files are real.)

**Records** (`scripts/data/build_orbit_cache.py`, one `.pt` per scene, loaded by `scripts/unified/eval_long_stream.py` and
`flc/unified/train_voxel_update.load_stream`): the whole video, every 3rd frame, 256 x 256 (`img_u8`), cameras (`c2w`, `K`),
VGGT-Omega 1B @512 depth and confidence from one joint pass over the run plus the scene's 11-frame clip, rescaled to the clip
scale (`depth`, `conf`, `scale_to_clip`), `held` (every 4th frame: never written to memory, the benchmark's targets), `turn`
(cumulative angle around the orbit centre, degrees) and `loop`. Test videos are uncut (up to ~110 kept frames); validation runs
are cut at 160 frames.

**Benchmark:** the observed frames are streamed in order in 11-frame windows; after every window the memory is rendered at the
held-out frames of the first orbit; all-pixel PSNR / LPIPS (holes count). `--loops 2` walks the video twice (6 orbits).
Commands: the READMEs of `09_voxel_update_f1_harm/` and `10_fusion_head_frozen_3k/`.

Derived from DL3DV-10K (frames resized; geometry from VGGT-Omega): for use within the team under DL3DV's terms; do not
redistribute. Built 2026-10-29 on the 8 x B200 machine.

sha256 orbit_eval_data.zip: df44e294a86f1c0d59501a73e2e21d502c911631396081ab3160ee057586483b