Download testset128/README.md from FrontRear/VLMdataset: direct link, hf CLI and curl.
- Browser
- Download file 6.15 kB
-
https://huggingface.co/datasets/FrontRear/VLMdataset/resolve/main/testset128/README.md
- Command line
-
hf download hf://datasets/FrontRear/VLMdataset/testset128/README.md
-
curl -L -o README.md https://huggingface.co/datasets/FrontRear/VLMdataset/resolve/main/testset128/README.md
testset128 — held-out ModularOffice, complete r6 release
128 main test video pairs, plus 12 separate saturated-color stress pairs. Each pair has49 frames at10FPS, native training RGB/conditions1280×704. Source camera A stays fixed; target B/C/D moves. A real animated UE character is present. This is a held-out test set: do not add its scene G085 / distribute7__ModularOffice to training.
Main quotas: insert19, remove19, move18, replace18, move+color18, replace+color18, move+replace+color18. The pairs share5 physical states,3 moving target tracks and8 appearance samples; they are not128 independent scenes/geometries. Stress12 is additional and excluded from the main128 metric count.
Download and extract
Download TESTSET128_ModularOffice_COMPLETE_r6.tar and its .sha256 from this folder. This is the entire current payload, not a patch requiring older releases.
sha256sum -c TESTSET128_ModularOffice_COMPLETE_r6.tar.sha256
tar -xf TESTSET128_ModularOffice_COMPLETE_r6.tar
sha256sum -c testset128/SHA256SUMS
Keep the two version folders next to one another as extracted:
testset128/
INFERENCE_INDEX_CORE128.json
GT_INDEX_CORE128.json
INFERENCE_INDEX_STRESS12.json
GT_INDEX_STRESS12.json
r5_conditions_complete_v1/ # immutable RGB, all geometry/camera/conditions, evaluation
r6_intensity_reference_compatible/ # current signed-intensity CSV/loader/acceptance
demos/
SHA256SUMS
All runtime paths are portable within this extracted tree. Absolute server paths in historical reports are provenance only. The archive contains regular file bytes, no external symlink/hardlink dependencies. Raw EXR, UE projects, models, caches and credentials are not included.
Conditions and entry points
INFERENCE_INDEX_CORE128.json is the current compact inference index and excludes target GT RGB. GT_INDEX_CORE128.json is a separate evaluation index. The corresponding STRESS12 indexes are separate. Complete physical PNG/video bindings and original parameter provenance remain in r5_conditions_complete_v1/TEST128_FOUR_CONDITIONS_MANIFEST_v2.json (lookup by pair_id).
Both ordinary PCF+bbox and depth-aware PCF+bbox are retained. The latter uses the same fixed world and camera rays to compare metric PCF/box depth, so nearer PCF geometry can occlude boxes and boxes can occlude farther geometry. Native lamp silhouette color/intensity and projected3D bbox color/intensity, camera matrices, native scene-depth visualizations, binary QA masks, wireframes and independently rendered fixture-reference images are included. Binary QA is not MaskColor.
Current intensity format (r6)
Use r6_intensity_reference_compatible/metadata_process.csv for the main128 lamp-silhouette interface, or metadata_process_bbox.csv for the bbox interface. _stress12.csv and _all140.csv are explicitly auxiliary alternatives. These full CSVs include a GT video field for offline validation; do not feed that field as an inference condition.
CSV alpha/gamma store raw target minus source, rounded to4 decimals. Full-precision GT coefficients and quantization error remain in the adapter provenance. Normalize alpha by0.9 and gamma by1.1 once at consumption. The r6 loader builds a3-channel signed field: inside the actual edit-region mask = relative_gamma, outside =−1. Unchanged intensity inside =0; negative deletion values remain negative. No additional2x−1 transform is applied to the signed scalar/tensor. Display PNGs are visualization only.
For removal, physical E0 target masks remain empty. A separate actual source-state footprint captured in the target camera localizes the negative edit. Lamp and bbox interfaces both use this rule. Earlier absolute gamma/1.1 grayscale masks and zero-background mathematical NPZ fields are preserved with distinct names; do not substitute them for the r6 consumer field.
HF reference CSV storage was observed from a real Validation_Real sample. The normalization/consumer arithmetic was verified against the supplied modified local LumiLens worktree with actual CPU Torch (including Root float32/bfloat16 boundary checks); the reference HF repository did not publish a loader. This release does not claim independent proof of that publisher's official loader revision. See R6_INTENSITY_README.md and the Root acceptance JSON.
RGB generation and evaluation
RGB is genuine Original factorized EXR synthesis, not generated/edit-model output. Capture uses24mm/10FPS,4000lm target components/socket, PT128 spatial×1temporal,8bounces,1000 clamp,native denoise, exposure−5, EXR inverse compensation32 and Original appearance sampling/fixed display. Completed1920×1080 captures are transformed by Original bilinear cover1280×720 then crop8top/bottom with matching K; no640 upscaling. The authorized r4 ambient revision and all historical changes are explicitly recorded in the included audits. Upload does not rewrite RGB, conditions, camera, illumination coefficients or exposure.
For metrics, install Python numpy/Pillow/opencv-python and system ffmpeg/ffprobe. Predictions must be49frames,1280×704,10FPS, named TEST128_001.mkv (or mp4) through TEST128_128. No metric-time resizing, realignment, exposure fitting or substitution is performed.
python testset128/r5_conditions_complete_v1/evaluation/evaluate_dataset.py \
--manifest testset128/r5_conditions_complete_v1/TEST128_FOUR_CONDITIONS_MANIFEST_v2.json \
--pred-dir ./predictions --encoding srgb_full_range \
--workers 2 --output ./test128_metrics.json
The evaluator reports sRGB/linear PSNR and SSIM, per-operation and equal-seven-class macro. Incomplete predictions do not receive a complete macro result. LPIPS weights/scoring are not included. GT-versus-itself reports are I/O checks, not model performance.
Preview
The demos show16 panels and all49 frames per selected operation. Their1920×1224 grids contain480×264 panels and are preview-only; original data stays1280×704. Purple/green extra examples are separate. Relative intensity display black=−1, midgray=0. Preview compression/crops do not replace lossless dataset data.