Reproduce “CornerNet: Detecting Objects as Paired Keypoints” 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
Hei Law, Jia Deng
CornerNet introduces an object detection framework that identifies objects as pairs of top-left and bottom-right corners, completely eliminating anchor boxes. This method achieved 42.2% AP on the MS COCO test-dev dataset, outperforming prior one-stage detectors and demonstrating superior bounding box localization with 23.4% AP at 0.9 IoU.
Key insights
- High-performance object detection is achievable without relying on anchor boxes, simplifying detector design and reducing hyperparameter complexity.
- Specialized pooling layers, like corner pooling, can effectively address the challenge of localizing keypoints that lack strong local visual evidence.
- The Hourglass network, originally designed for human pose estimation, serves as an effective backbone for object detection when combined with an appropriate keypoint-based formulation and loss functions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.