CylinderDepth

CVPR 2026 Workshops (3DMV)

CylinderDepth: Cylindrical Spatial Attention for Multi-View Consistent Self-Supervised Surround Depth Estimation

Project Page | Paper | Supplementary Material

Data Preparation

To generate the ground-truth depth labels:

python tools/export_gt_depth_ddad.py
python tools/export_gt_depth_nusc.py

To generate the overlap ground-truth depth labels:

python tools/export_overlap_depth_ddad.py
python tools/export_overlap_depth_nuscenes.py

Checkpoints

You can download the pre-trained checkpoints for DDAD and nuScenes here: https://huggingface.co/samerabualhanud/CylinderDepth/tree/main

Environment Setup

 cd CylinderDepth
 conda env create -f CylinderDepth.yml
 conda activate CylinderDepth
 cd external/dgp
 pip install -r requirements.txt -r requirements-dev.txt
 pip install --editable .

Training

To train the model:

Change the data paths in dataset/ddad_dataset.py,dataset/nusc_dataset.py and the config files.

DDAD

python  train.py  \

--config_file  ./configs/ddp/nuscenes/nusc_baseline_352_ddp_min_1.0_front_sp_con_0.001_sptp_con_0.05_flipv5.yaml

nuScenes

python  train.py  \

--config_file  ./configs/ddp/baseline_ddp_384_front_sp_con_pre_0.001_sptp_con_0.2_flipv5.yaml

Evaluation

To evaluate the model on the overlap depth, set overlap in the config file to True.

To evaluate the model:

python  eval.py  \

--config_file  ./configs/ddp/baseline_ddp_384_front_sp_con_pre_0.001_sptp_con_0.2_flipv5.yaml  \

--weight_path  ./results/baseline_ddp_384_front_sp_con_pre_0.001_sptp_con_0.2_flipv5/models/weights_19

Cite

If you find our work useful, please cite our paper:

@inproceedings{abualhanud2026cylinderdepth,
  title={CylinderDepth: Cylindrical Spatial Attention for Multi-View Consistent Self-Supervised Surround Depth Estimation},
  author={Abualhanud, Samer and Grannemann, Christian and Mehltretter, Max},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={169--179},
  year={2026}
}

Notes

We would like to thank the authors of VFDepth, SurroundDepth, CVCDepth, and MonoDepth2. This codebase builds upon and benefits greatly from their valuable open-source contributions.