This is a PyTorch implementation of the OFTNet network from the paper Orthographic Feature Transform for Monocular 3D Object Detection. The code currently supports training the network from scratch on the KITTI dataset - intermediate results can be visualised using Tensorboard. The current version of the code is intended primarily as a reference, and for now does not support decoding the network outputs into bounding boxes via non-maximum suppression. This will be added in a future update. Note also that there are some slight implementation differences from the original code used in the paper. Please see train.py for details of training options.

If you find this work useful please cite the paper using the citation below.

  title={Orthographic feature transform for monocular 3d object detection},  
  author={Roddick, Thomas and Kendall, Alex and Cipolla, Roberto},  
  journal={British Machine Vision Conference},