Sequential Multi-View Fusion Network for Fast LiDAR Point Motion Estimation

Official code for SMVF

Sequential Multi-View Fusion Network for Fast LiDAR Point Motion Estimation, Gang Zhang, Xiaoyan Li, Zhenhua Wang. (https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136820282.pdf) Accepted by ECCV2022

NEWS

• [2022-07-03] SMVF is accepted by ECCV 2022
• [2022-03-07] SMVF achieves 1st place in SemanticKITTI Moving Object Segmentation leaderboard.

1 Dependency

CUDA>=10.1
Pytorch>=1.5.1
PyYAML@5.4.1
scipy@1.3.1

2 Training Process

2.1 Installation

cd deep_point
python setup.py install

2.2 Prepare Dataset

Please download the SemanticKITTI dataset to the folder SemanticKITTI and the structure of the folder should look like:

./
├── 
├── ...
└── dataset/
    ├──sequences
        ├── 00/         
        │   ├── velodyne/
        |   |	├── 000000.bin
        |   |	├── 000001.bin
        |   |	└── ...
        │   └── labels/ 
        |       ├── 000000.label
        |       ├── 000001.label
        |       └── ...
        ├── 08/ # for validation
        ├── 11/ # 11-21 for testing
        └── 21/
	        └── ...

And download the object bank on the SemanticKITTI to the folder object_bank_semkitti and the structure of the folder should look like:

./
├── bicycle
├── bicyclist
├── car
├── motorcycle
├── motorcyclist
├── other-vehicle
├── person
├── truck

2.3 Training Script

python3 -m torch.distributed.launch --nproc_per_node=8 train.py --config config/config_smvf_sgd_ohem_vfe_k2_fp16_48epoch.py

3 Evaluate Process

python3 -m torch.distributed.launch --nproc_per_node=8 evaluate.py --config config/config_smvf_sgd_ohem_vfe_k2_fp16_48epoch.py --start_epoch 0 --end_epoch 47