• CUDA >= 9.0
• Python 3
• pyquaternion, Matplotlib, PIL, numpy, cv2, tqdm (and other relevant packages which can be easily installed with pip or conda)
• PyTorch >= 1.1
  • Note that, the MGDA-related code currently can only run on PyTorch 1.1 (due to the official implementation of MGDA). Such codes include min_norm_solvers.py, train_single_seq_MGDA.py and train_multi_seq_MGDA.py.

1. To run the code, first need to add the path to the root folder. For example:

export PYTHONPATH=/home/pwu/PycharmProjects/MotionNet:$PYTHONPATH
export PYTHONPATH=/home/pwu/PycharmProjects/MotionNet/nuscenes-devkit/python-sdk:$PYTHONPATH

2. Data preparation (suppose we are now at the folder MotionNet):
   • Download the nuScenes data.
   • Run command python data/gen_data.py --root /path/to/nuScenes/data/ --split train --savepath /path/to/the/directory/for/storing/the/preprocessed/data/. This will generate preprocessed training data. Similarly we can prepare the validation and test data.
   • See readme.md in the data folder for more details.
3. Suppose the generated preprocessed data are in folder /data/nuScenes_preprocessed, then:
   • To train the model trained with spatio-temporal losses: python train_multi_seq.py --data /data/nuScenes_preprocessed --batch 8 --nepoch 45 --nworker 4 --use_bg_tc --reg_weight_bg_tc 0.1 --use_fg_tc --reg_weight_fg_tc 2.5 --use_sc --reg_weight_sc 15.0 --log. This command will train the model with spatio-temporal consistency losses. See the code for more details.
   • To train the model with MGDA framework: python train_multi_seq_MGDA.py --data /data/nuScenes_preprocessed --batch 8 --nepoch 70 --nworker 4 --use_bg_tc --reg_weight_bg_tc 0.1 --use_fg_tc --reg_weight_fg_tc 2.5 --use_sc --reg_weight_sc 15.0 --reg_weight_cls 2.0 --log.
   • The pre-trained model for train_multi_seq.py can be downloaded from Google Drive or Dropbox
   • The pre-trained model for train_multi_seq_MGDA.py can be downloaded from Google Drive or Dropbox
   • The files train_single_seq.py and train_single_seq_MGDA.py train MotionNet exactly in the same manner, except without utilizing spatio-temporal consistency losses.
4. After obtaining the trained model, e.g., model.pth for train_multi_seq.py, we can evaluate the model performance as follows:
   • Run python eval.py --data /path/to/the/generated/test/data --model model.pth --split test --log . --bs 1 --net MotionNet. This will test the performance of MotionNet.

To visualize the results:

1. Generate the predicted results into .png images: run python plots.py --data /path/to/nuScenes/data/ --version v1.0-trainval --modelpath model.pth --net MotionNet --nframe 10 --savepath logs
2. Assemble the generated .png images into .gif or .mp4: python plots.py --data /path/to/nuScenes/data/ --version v1.0-trainval --modelpath model.pth --net MotionNet --nframe 10 --savepath logs --video --format gif

• train_single_seq.py and train_single_seq_MGDA.py train the model without using spatio-temporal consistency losses.
• train_multi_seq.py and train_multi_seq_MGDA.py train the model with spatio-temporal losses.
• model.py contains the definition of MotionNet.
• min_norm_solvers.py is for multi-objective optimization framework (MGDA).
• eval.py contains several metrics for evaluating the model performance.
• plots.py contains utilities for generating the predicted images/videos.
• data/data_utils.py includes the utility functions for preprocessing the nuScenes data.
• data/gen_data.py generates the preprocessed nuScenes data.
• data/nuscenes_dataloader.py the dataloader for model training/validation/testing.
• nuscenes-devkit this folder is based on the nuScenes official API, and modified to include many other utilities.

@inproceedings{wu2020motionnet,
  title={MotionNet: Joint Perception and Motion Prediction for Autonomous Driving Based on Bird's Eye View Maps},
  author={Wu, Pengxiang and Chen, Siheng and Metaxas, Dimitris},
  booktitle={CVPR},
  year={2020}
}