DA-W: Weather-Conditioned Depth Anything

Checkpoints for DA-W, a weather-robust monocular depth estimation model built on Depth Anything V2 for zero-shot relative depth under fog, rain, snow and low light (ECCV 2026).

A Style Filter extracts a content-independent 64-D weather embedding from Gram-matrix statistics. The embedding is injected into the frozen Depth Anything V2 ViT-S backbone through zero-initialized AdaLN-Zero heads in the DPT decoder, so the model starts exactly at the pretrained baseline and adapts to weather without forgetting clean-scene performance.

• Project page: https://zhaoming-tamu.github.io/WCDA
• Code: https://github.com/taco-group/DA-W

Files

File	Description
daw_vits_stage2.pth	DA-W (ViT-S) depth model with AdaLN weather conditioning (Stage II)
daw_style_filter_stage1.pth	Style Filter producing the 64-D weather embedding (Stage I)

Both files are plain PyTorch state_dicts and must be used together with the code repository.

Usage

pip install -r requirements.txt
hf download qgfvadfuvads/DA-W --local-dir checkpoints
python infer.py --img-dir /path/to/images --save-dir runs/wild

Results

Zero-shot relative depth (AbsRel ↓ / δ₁ ↑), ViT-S encoder:

Method	NuScenes-night	RobotCar-night	DS-rain	DS-cloud	DS-fog	KITTI-C Dark	KITTI-C Snow	KITTI-C Fog
Depth Anything V2	0.200 / 0.725	0.239 / 0.518	0.125 / 0.840	0.151 / 0.798	0.103 / 0.890	0.130 / 0.832	0.115 / 0.872	0.097 / 0.905
DA-W	0.194 / 0.737	0.239 / 0.513	0.123 / 0.842	0.151 / 0.795	0.101 / 0.896	0.126 / 0.837	0.107 / 0.884	0.093 / 0.910

License

Released under CC BY-NC 4.0 for non-commercial research use only. The model is initialized from Depth Anything V2 ViT-S (Apache-2.0) and distilled with the Depth Anything V2 ViT-L teacher (CC-BY-NC-4.0); the code repository additionally builds on DepthAnything-AC (CC BY-NC 4.0).

Citation

@inproceedings{xu2026weather,
  title     = {Weather-Conditioned Depth Anything},
  author    = {Xu, Zhaoming and Hu, Chan-Wei and Huang, Kuan-Ru and
               Zhu, Zihao and Li, Renjie and Zhou, Yang and Tu, Zhengzhong},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}