metadata
pretty_name: WildShadow-Image
license: other
license_name: mixed-source-dataset
tags:
- image
- shadow-removal
- arxiv:2607.26203
WildShadow-Image
WildShadow-Image is the paired image split used by the image-mode experiments of WildShadowRemover.
- Code: ylxclimax/WildShadowRemover
- Paper: arXiv:2607.26203
| Split | INS | iCity | Infinigen | HSSD | Total |
|---|---|---|---|---|---|
| Train | 30,000 | 3,000 | 3,000 | 0 | 36,000 pairs |
| Test | 2,000 | 800 | 400 | 300 | 3,500 pairs |
HSSD is evaluation-only.
Layout
WildShadow-Image/
train/ins/{origin,shadow_free}/shard_000...shard_005/
train/{icity,infinigen}/{origin,shadow_free}/
test/{ins,hssd,icity,infinigen}/{origin,shadow_free}/
manifests/pairs.csv
BUILD_INFO.json
The same filename in origin/ and shadow_free/ forms one pair. Each dataset
root can be passed directly to the image training launcher in
WildShadowRemover with single-image mode enabled.
Depth is not included because it is a deterministic model-derived condition, not source ground truth. Generate current float16 DA3 depth with:
MODE=image ROOTS=/path/to/WildShadow-Image/train/ins,\
/path/to/WildShadow-Image/train/icity,\
/path/to/WildShadow-Image/train/infinigen \
bash scripts/prepare_depth.sh
This prevents legacy u8/u16 visualization depth from being mixed with the current floating-point training contract.
Licensing
This is a mixed-source dataset. See LICENSE and SOURCE_LICENSES.md; the
archive does not grant rights beyond the terms of the underlying sources.