awwkl/zwm-stereo-depth-eval

ZWM stereo-depth eval dataset

Stereo image pairs with two annotated points per pair (with ground-truth depth) used to evaluate depth perception in the ZWM repo. The eval reuses ZWM's counterfactual flow engine: predict EPE for each point's "motion" between stereo halves, then order depths by EPE magnitude (closer = larger EPE).

Layout

stereo_depth/
├── dataset.json                # 328 entries (164 pairs across kitti_1000 + kitti_500_flipud)
└── frames/
    ├── cropped_without_crosses/    # query frames (model input)
    ├── cropped_stereo_mate/        # target frames (model input)
    └── cropped_with_crosses/       # annotated viz overlays (not consumed by the eval)

query_frame_file / target_frame_file / cropped_with_crosses_file in the JSON are paths relative to stereo_depth/frames/.

Usage with the ZWM repo

huggingface-cli download awwkl/zwm-stereo-depth-eval --repo-type dataset --local-dir data/evals/depth/
bash scripts/eval/depth/eval_stereo_depth.sh
bash scripts/eval/depth/grade_stereo_depth.sh

Source

Stereo crops + point annotations from the UniQA-3D stereo benchmark (KITTI-1000 and KITTI-500-flipud subsets, ~164 image pairs total).