CCRss/oof-vlm-judge-k16

OOF VLM-judge package (K=16, cut 50%)

Companion package for "Present but Unidentifiable: Out-of-Frame Completion in Single-Image 3D Generators" (WACV). 70 real truncated objects x 16 sampled completions each, for selector-with-world-knowledge (VLM-judge) experiments vs the paper's Table 6 blind selectors.

Access is gated: request access and the author will approve.

Format — oof_vlm_package.tar (2.6 GB)

oof_vlm_package/
├── README.md                    # full generation settings & cautions
└── objNNN/                      # one dir per object, obj000 … obj069 (70 objects, cut = 50%)
    ├── s00.npz … s15.npz        # the 16 sampled completions (geometry)
    ├── sSS_az{0,1,2,3}.png      # 4 rendered views per sample, 448x448 (az0 = front)
    ├── objNNN_sheet.png         # 4x4 contact sheet of all 16 samples (front view, cd in title)
    └── scores.json              # per-sample Chamfer + generation settings

sSS.npz — one sampled 3DGS completion (200,000 Gaussians):

key	shape	dtype	meaning
xyz	(200000, 3)	float16	Gaussian centers
opacity	(200000,)	float16	opacities
scale	(200000, 3)	float16	per-axis scales
rgb	(200000, 3)	uint8	colors

scores.json per object:

• cd_regen — list of 16 floats: Chamfer of sample s vs pseudo-GT. Authoritative for these saved geometries (index = sample = seed).
• cd_reference — the paper's originally stored Chamfer for the same (object, seed); provenance check, matches cd_regen to ~2% (occasional single-seed mode flips expected, rank order stable).
• abs_diff, mean_abs_diff — the comparison above.
• obj, keep (=0.5 -> cut 50%), steps (16), gs (4.0), shift (3.0), numg (200000), seed_convention.

Conventions: sample s generated with seed s (fully reproducible; frozen TripoSplat ckpts); pseudo-GT = full-image reconstruction with seed 0; renders are depth-shaded geometry (no texture bias for the judge).

Cautions for the VLM-judge setup

1. Position bias: randomize sample order/positions in any grid shown to the VLM and average over permutations.
2. Coverage gaming: do not let the judge select on visible-region fidelity alone — the question is the hidden side.

Conditioning inputs

inputs_cut50.zip (9.7 MB) — the exact conditioning images the model saw: objNNN_input_cut50.png = the object image with everything below y0 + (y1-y0)*0.5 of the object's vertical bbox set to black (bottom-50% truncation), where (y0,y1) is the foreground extent (pixel-sum > 10). The same files are also included per-object inside the tar as objNNN/input_cut50.png for future repacks. Full untruncated sources are in issai/oof-real-ood-preview (object.png).