Object dynamics prediction with attention
To run the code, first run preprocess.py to convert text file annotation to maps. Changes need to be made to load box dimensions correctly.
After data is preprocessed, run train_nn.py, which has two training modes: batch and recurrent. Batch training always feeds in the ground truth and predicts the next frame. Recurrent training gives a certain number of ground truth frames (seen_step) and predicts the next couple of frames (fut_step). Finally, the model can be evaluated using mode_eval.
Batch training works fine now, but train_recurrent and mode_eval might need some debugging. Finally, we also need some code to visualize the output to better compare against the ground truth.