Reproduce “DeTra: A Unified Model for Object Detection and Trajectory Forecasting” 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
Sergio Casas, Ben Agro, Jiageng Mao, Thomas Gilles, Alexander Cui, Thomas Li, Raquel Urtasun
Researchers from Waabi and the University of Toronto developed DeTra, a unified model that re-frames object detection and trajectory forecasting as a single trajectory refinement problem directly from LiDAR and HD maps. The model achieves state-of-the-art performance on Argoverse 2 Sensor and Waymo Open Dataset, improving joint detection and forecasting metrics (OccAP, TrajAP) by up to 13.5% over previous methods.
Key insights
- Object detection can be effectively re-framed as the initial pose of an object's trajectory, enabling a truly unified and jointly optimized perception and prediction system.
- Iterative refinement using Transformer blocks significantly improves both detection accuracy and the plausibility and diversity of multi-modal trajectory forecasts.
- Efficient and specialized attention mechanisms (local deformable LiDAR, k-NN map, and factorized self-attention) are crucial for handling multi-modal, temporal, and multi-object scene reasoning within a unified architecture.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.