CUDA 11.1
pytorch 1.9.0
The following is required to be installed before running:
torvision 0.10.0
pyyaml
scipy
matplotlib
opencv-python
tensorboardX
Run the following commands to install pointnet2 and weighted FPS:
cd external
cd pointnet2
python setup.py install
cd ..
cd weighted_FPS
python setup.py install
Download TOS dataset with bbox, instance label and semantic label. Then, modify the dataset_dir and save_dir under ./utils/generate_gt_heatmap.py.
Then run this script to generate the ground truth objectness map that is mentioned in the paper.
Then, put the TO-xxx-wHM folder under ./data
- Train
For training the TO-crowd or TO-vanilla, it requires pretraining the heatmap module firstly, to train the heatmap module, run the following commands:
python main.py --mode train --config ./configs/train_heatmap.yaml
Then, modify the hm_pretrain_path in ./configs/train_vote_adaptive_desk.yaml to point to the pretrain heatmap module's weight.
Finally run the adaptive votenet by:
python main.py --mode train --config ./configs/train_vote_adaptive_desk.yaml
If you wish to run original VoteNet on our dataset, you may refer to the usage on their official repo and our implementation here.