Reproduce “RESA: Recurrent Feature-Shift Aggregator for Lane Detection” 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
Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, Deng Cai
Researchers from Zhejiang University and Fabu Inc. introduced RESA, a recurrent feature-shift aggregator, and BUSD, a bilateral up-sampling decoder, to enhance lane detection. The proposed architecture achieved an F1-measure of 75.3% on CULane and 96.82% accuracy on Tusimple, while operating at a real-time speed of 28 ms.
Key insights
- Recurrently shifting and aggregating feature maps with dynamic strides enables efficient global spatial information gathering with logarithmic complexity, overcoming the limitations of sequential spatial convolution.
- The bilateral design of the decoder effectively combines coarse and fine-grained features during up-sampling, significantly improving the fidelity of detail recovery for thin and long structures like lanes.
- The proposed RESA and BUSD modules are complementary, enhancing both the robustness of feature extraction under challenging conditions and the precision of pixel-wise lane boundary predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.