In configs/local_nusc.yaml

• base_dir: Put official nuScenes here, e.g. {base_dir}/v1.0-trainval
• image_dir: Put segmentation results here. We use Mask2Former to segment source images. Folder structure is like {image_dir}/{sweeps/samples}/seg_CAM_FRONT. We provide processed data on google drive, including all semantic images used in this paper. "Scene-1" and "Scene-2" of the paper contain scene-0063, scene-0064, scene-0200, scene-0283 and scene-0109, scene-0508, scene-0523, scene-0821 respectively.

In configs/local_kitti.yaml

• base_dir: Put official kitti odometry dataset here, e.g. {base_dir}/sequences
• image_dir: Put segmentation results here. Also we use Mask2Former and folder structure is like {image_dir}/seg_sequences. We provide processed data on google drive, including all semantic images and poses we used.

torch==1.10.2+cu111
torchvision==0.11.3+cu111
torchaudio==0.10.2+cu111
pytorch3d==0.6.1
pymeshlab==2021.10 
scipy opencv-py thon tqdm wandb python3.8

For wandb usage, please visit here.

• Modify configs/local_nusc.yaml

  • change wandb configs
  • change base_dir and image_dir according to your folder
  • change clip_list to train one scene or multiple scenes.

• Modify wandb url and api_key in run_local.sh and then run sh run_local.sh to start training.

• Modify configs/local_kitti.yaml

  • change wandb configs
  • change base_dir and image_dir according to your folder
  • change sequence to choose which sequence to train.
  • modify choose_point and bev_x/y_length to choose which sub area to train.

• Modify wandb url and api_key in run_local.sh and then run sh run_local.sh to start training.

• Modify configs/nusc_eval.yaml
  • change model_pathand pose_path where your trained models saved.
  • Make sure other training parameters are same as your configs when training.
  • This is a simple evaluation script and only support batch_size: 1

• This is the first version of RoMe, and it is hard to reconstruct steep slopes.
• Using SfM(structure from motion) or MVS(multi-view stereo) points and lidar points will give strong supervision.
• Monodepth estimation would be useful like monosdf.
• We are trying to use other mesh render methods like nvdiffrec.

@misc{mei2023rome,
      title={RoMe: Towards Large Scale Road Surface Reconstruction via Mesh Representation}, 
      author={Ruohong Mei and Wei Sui and Jiaxin Zhang and Xue Qin and Gang Wang and Tao Peng and Cong Yang},
      year={2023},
      eprint={2306.11368},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}