Enhancing Glass Surface Reconstruction via Depth Prior for Robot Navigation

Paper | GlassRecon Dataset

Dataset Structure

The dataset is organized as follows:

GlassRecon/
├── images/                  # RGB images (PNG)
├── intrinsics/              # JSON files with camera intrinsics & depth scale
├── masks/                   # Binary masks for glass regions (PNG)
├── sensor_depths/           # Raw sensor depth maps (PNG)
├── completed_depths/        # Completed depth maps (NPY)
└── evaluation_depths/
    ├── depths_npy/          # Filtered depth maps (NPY) – glass regions that could not be completed are masked out
    ├── depths_vis/          # Visualizations (PNG)
    └── pointclouds/         # 3D point clouds (PLY) - back-projected from depths_npy using intrinsics

Evaluation

Use eval.py to compute metrics (AbsRel, δ < 1.25) between predicted depth maps and ground truth.

python eval.py --image-folder IMAGE_PATH \
               --pred-folder PREDICTED_DEPTH_PATH \
               --sensor-depth-folder SENSOR_DEPTH_PATH \
               --gt-depth-folder GT_DEPTH_PATH \
               --depth-scale DEPTH_SCALE \
               --outdir OUTPUT_PATH

If the predicted depth maps represent inverse depth, add the --inverse-depth flag:

python eval.py ... --inverse-depth

Citation

If you find out work useful please cite

@article{zheng2026enhancing,
  title={Enhancing Glass Surface Reconstruction via Depth Prior for Robot Navigation},
  author={Zheng, Jiamin and Yu, Jingwen and Chen, Guangcheng and Zhang, Hong},
  journal={arXiv preprint arXiv:2604.18336},
  year={2026}
}