Deep3DBox's MXNet implementation.
- Train Faster RCNN on KITTI to gain 2D BBox predictions.
- Add Jitter to GT BBox, then use it and the above generated bbox to train Deep3DBox.
- inherit IMDB to create KITTI data iter, provide_data and provide_label
- load pre-trained res-next101, train the model.
- Use cpp devkit to evaluate the model, use MATLAB to visualize.