Report | [Video](add link for video)
In this project we focus on Bird's Eye View (BEV) prediction based on monocular photos taken by the cameras on top of the car. We experiment with Determinisitic autoencoders, stochastic variational autoencoders, generative adversarial networks for generating Bird's eye view road layout and Bird's eye view of vehicles on the road indirectly. THe best performing models on the training set use GANs whereas the maximum test performance was from the deterministic model. Our models achieve 0.904 val threat score on the road layout prediction task and 0.044 val threat score on the BB prediction task.
Use generate_labels.py
to generate
- vehicles mask
- road mask
- warped and glued photos
Refer to src/
for code used to train and test road layout prediction models.
- GANs
src/GANmodels
- Deterministic models and Retinanet
src/SupModels
- training and validation scripts
src/trainer
- training and validation scripts
src/trainer
Implemented PIRL and SIMCLR SSL techniques in src/SSLmodels.py
Libraries used
- simclr https://arxiv.org/abs/2002.05709
- PIRL https://arxiv.org/abs/1912.01991
- retinanet https://arxiv.org/abs/1708.02002
- rotation based object detection https://arxiv.org/pdf/1911.08299.pdf