Neural Networks (LTAT.02.001)

Motion prediction model for self driving

Download dataset

The dataset has to be downloaded inside "input/" folder in the root location of the repo. Download the dataset from here

input/
  +- scenes/
        +- sample.zarr
        +- train.zarr
        +- validate.zarr
        +- test.zarr
  +- aerial_map/
        +- aerial_map.png
  +- semantic_map/
        +- semantic_map.pb
  +- meta.json

Install Dependencies

pip install l5kit
pip install timm

Training

Following backbone CNNs have been added with their respective variants

• Resnets
• Mixnets

1- Open notebook_train.ipynb and follow step by step instructions to train either unimodal or multimodal model.

2- After training use notebook_test.ipynb to evaluate your model, like the training notebook all the instructions are added in this file as well.

Results

References

1- Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., ... & Ondruska, P. (2020). One thousand and one hours: Self-driving motion prediction dataset. arXiv preprint arXiv:2006.14480.

2- Targ, Sasha, Diogo Almeida, and Kevin Lyman. "Resnet in resnet: Generalizing residual architectures." arXiv preprint arXiv:1603.08029 (2016).

3- Tan, Mingxing, and Quoc V. Le. "Mixconv: Mixed depthwise convolutional kernels." arXiv preprint arXiv:1907.09595 (2019).