Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
label
class label
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
03dfront
End of preview. Expand in Data Studio

WildShadow-Video

WildShadow-Video is the paired video split used for the video experiments in WildShadowRemover. Each clip contains temporally ordered, frame-aligned shadow and shadow-free RGB videos.

Split 3D-FRONT iCity Infinigen HSSD Total
Train 4,500 900 700 0 6,100 clips
Test 400 100 100 50 650 clips

HSSD is evaluation-only. Test clips are never augmented. Train and test source sequences are disjoint and are also checked using content fingerprints.

Layout

WildShadow-Video/
  train/{3dfront,icity,infinigen}/
  test/{3dfront,icity,infinigen,hssd}/
  manifests/clips.csv
  scripts/extract_frames.py
  scripts/extract_dataset.py
  scripts/qc_frame_pairs.py
  scripts/validate_release.py
  SHA256SUMS

Each clip has two files:

<clip_id>__shadow.mp4
<clip_id>__shadow_free.mp4

Videos use lossless H.264 RGB (libx264rgb, CRF 0, RGB24) at 30 FPS. FPS is a portable playback convention; frame order is the training signal. Decoding to RGB reproduces the source RGB pixels exactly. Source alpha channels, where present, are not included because the model consumes RGB. The build audit checks all video frame counts and dimensions and performs exact source/decoded RGB comparison on 50 deterministic clips from every split/dataset group.

Extract a pair back into numbered PNG frames with:

python scripts/extract_frames.py \
  --shadow train/icity/example__shadow.mp4 \
  --shadow-free train/icity/example__shadow_free.mp4 \
  --output /path/to/example

Extract a complete split with:

python scripts/extract_dataset.py --root . --split train \
  --output /path/to/frame_dataset --workers 8

Depth is intentionally not stored in lossy image/video form. Generate float depth from the extracted origin frames using the Depth Anything 3 preparation script distributed with WildShadowRemover. Stage-II masks and latents are model artifacts and are likewise not part of this source dataset archive.

Verify downloaded files with:

sha256sum -c SHA256SUMS
python scripts/validate_release.py --root .

Licensing

See SOURCE_LICENSES.md. This archive does not replace or broaden the licenses of the source datasets and assets.

Downloads last month
86

Paper for dcowg/wildshadow-video