Reproduce “DAMO-StreamNet: Optimizing Streaming Perception in Autonomous Driving” 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
Jun-Yan He, Zhi-Qi Cheng, Chenyang Li, Wangmeng Xiang, Binghui Chen, Bin Luo, Yifeng Geng, Xuansong Xie
DAMO-StreamNet, from DAMO Academy and Carnegie Mellon University, presents a real-time perception framework for autonomous driving, incorporating a dynamic receptive field FPN and asymmetric knowledge distillation. The framework achieves 37.8% sAP on the Argoverse-HD dataset, demonstrating improved object detection across scales and enhanced long-term forecasting capabilities for continuous video streams.
Key insights
- DRFPN consistently enhances feature aggregation and temporal feature alignment, providing adaptive receptive fields crucial for detecting objects across varying scales.
- Utilizing longer-term temporal information (up to N=3 frames) improves future prediction accuracy, though an optimal balance is necessary to avoid performance degradation.
- Logits-based knowledge distillation effectively transfers 'future knowledge' to the current prediction, particularly benefiting models with shorter temporal modeling capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.