Reproduce “MPA: MultiPath++ Based Architecture for Motion Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stepan Konev
MPA, a MultiPath++ based architecture, placed 3rd in the Waymo Motion Prediction Challenge 2022 by refining model architecture, training strategies, and post-processing for diverse agent types. The model achieved a Total Soft mAP of 0.3930 and a Total mAP of 0.3866 on the challenge test set.
Key insights
- Optimal model architecture varies by agent type; a single decoder worked best for vehicles and pedestrians, while multiple decoders with attention performed better for cyclists.
- Non-Maximum Suppression (NMS) in post-processing demonstrably improves prediction diversity and accuracy, leading to higher Soft mAP scores.
- Data augmentation, specifically random masking of historical data, is beneficial for preventing overfitting, particularly for less represented agent classes like pedestrians and cyclists.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.