Unofficial proof-of-concept implementation of
Scene Transformer: A unified architecture for predicting multiple agent trajectories (https://arxiv.org/abs/2106.08417)
paper by Google Brain.
We implement their model and apply to SDC dataset. Only factorized self-attention part of the model is implemented, which is responsible for trajectory prediction based on time and agents in a scene. Cross-attetion part accountable for road graph is missing.
The structure of repo is the following:
scenetransformer.py: model impltemetation,sdcdataset.py: dataloader implementation for SDC datasetsdc_train_loop.py: train loop + loss implementation