Reproduce “Deep Snake for Real-Time Instance Segmentation” 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

Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, Xiaowei Zhou

Deep Snake, developed by Zhejiang University and Deepwise AI Lab, introduces a learning-based active contour model for real-time instance segmentation, leveraging novel circular convolution and a two-stage pipeline. The method achieves state-of-the-art performance on Cityscapes with 4.6 fps, outperforming PANet by 0.9 AP, and demonstrates strong results across various datasets with real-time inference speeds.

Key insights

  • Tailoring network architecture with circular convolution to the contour's cycle-graph topology significantly enhances feature learning and deformation capabilities over generic graph convolutions.
  • A robust two-stage initialization (diamond to octagon via extreme points) combined with iterative deformation effectively compensates for initial object localization errors from detectors and refines contours precisely.
  • The compact contour representation, optimized learning-based deformation, and efficient pipeline enable real-time instance segmentation competitive with or superior to pixel-based methods in both speed and accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.