Reproduce “PIDNet: A Real-time Semantic Segmentation Network Inspired by PID Controllers” 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
Jiacong Xu, Zixiang Xiong, Shankar P. Bhattacharyya
PIDNet introduces a real-time semantic segmentation network inspired by Proportional-Integral-Derivative (PID) controllers, featuring a novel three-branch architecture to overcome the detail loss in existing two-branch designs. It achieves a new state-of-the-art trade-off between inference speed and accuracy on benchmarks like Cityscapes, with PIDNet-L reaching 80.6% mIoU at 31.1 FPS.
Key insights
- Two-branch semantic segmentation networks behave analogously to Proportional-Integral (PI) controllers, leading to an 'overshoot' problem where fine details are lost during feature fusion.
- Adding a 'Derivative' component via a dedicated branch for high-frequency features can 'dampen' this overshoot, effectively preserving boundary information and fine details.
- The synergistic interaction of modules like Pag (selective learning) and Bag (boundary-attention-guided fusion) is crucial for the PIDNet architecture's ability to maintain both global context and local detail.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.