2nd Solution for Urban3D@ICCV2021

This repository is for the runner up solution for the challenge Urban3D@ICCV2021.

News

• We achieve the 3rd place in Urban3D@ECCV2022！The 1st solution also uses this codebase!

Getting Started

Requirements

• Ubuntu 16.04
• Anaconda with python=3.7
• tensorFlow=1.14
• cuda=10.1
• cudnn=7.6.5
• others: pip install termcolor opencv-python toposort h5py easydict

Compile custom operators

sh init.sh

Data preparation

(1) Download the files named "data_release.zip" here. Uncompress it.

(2) Preparing the dataset

python datasets/prepare_data.py --dataset_path $YOURPATH
cd $YOURPATH; 
cd ../; mkdir original_block_ply; mv data_release/train/* original_block_ply; mv data_release/test/* original_block_ply;
mv data_release/grid* ./

(3) The data should organized in the following format:

/Dataset/SensatUrban/
          └── original_block_ply/
                  ├── birmingham_block_0.ply
                  ├── birmingham_block_1.ply 
		  ...
	    	  └── cambridge_block_34.ply 
          └── grid_0.200/
	     	  ├── birmingham_block_0_KDTree.pkl
                  ├── birmingham_block_0.ply
		  ├── birmingham_block_0_proj.pkl 
		  ...
	    	  └── cambridge_block_34.ply 

Training

We train our model with two V100 GPUs. If you want to use other type of GPU with smaller batch size or other model architecture, you can modify the configuration in cfgs/sensaturban/baseline_version1.yaml.

python function/train_evaluate_sensaturban.py --cfg cfgs/sensaturban/baseline_version1.yaml  --gpus 0 1

• The results will be saved in ./log/.
• You can use --trainval to train on both trianing and validation sets.

Evaluation

python function/evaluate_sensaturban.py  --load_path [YOUR_MODEL_PATH]  --cfg cfgs/sensaturban/baseline_version1.yaml

• The results will be saved in ./log_eval/.

Submit to benchmark

python function/test_sensaturban.py  --load_path [YOUR_MODEL_PATH]  --cfg cfgs/sensaturban/baseline_version1.yaml

• The results will be saved in ./log_test/.
• Submit the results to the evaluation server.

Pre-trained model

We currently release a checkpoints with about 66% mIoU, and more powerful checkpoints will be released in the future.

Input	mIoU	Accuracy	Checkpoints
xyz+rgb	65.8	94.0	BaiduNetdisk:1wt2

You can specific the checkpoint in evaluation or testing by using

--load_path [CHECKPOINTS_PATH]/model.ckpt-600 

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{yan2021urban3d,
  title={Urban3D-2021-2rd},
  author={Xu Yan},
  booktitle={https://github.com/yanx27/Urban3D-2021-2nd},
  year={2021}
}

Acknowledgement

This project is not possible without multiple great opensourced codebases.

• SensatUrban
• CloserLook3D