Sparse R-CNN: An End-to-End Framework for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection is a fundamental computer vision task.
• Current methods often rely on dense, pre-defined object candidates like anchor boxes.
• This dense prior approach requires complex design and post-processing steps.

Purpose of the Study:

• To present Sparse R-CNN, a simplified and sparse object detection method.
• To eliminate the need for hand-designed object candidates and complex label assignment.
• To establish an end-to-end object detection framework.

Main Methods:

• Replaced dense, numerous anchor boxes with a fixed, small set of learned object proposals (N).
• Utilized these sparse proposals for object classification and localization.
• Removed the non-maximum suppression (NMS) post-processing step.

Main Results:

• Sparse R-CNN achieves highly competitive accuracy and runtime performance.
• Demonstrates strong training convergence compared to established baselines.
• Successfully validated on challenging COCO and CrowdHuman datasets.

Conclusions:

• Sparse R-CNN offers a simpler, more efficient object detection paradigm.
• The method inspires rethinking dense priors in object detection.
• Paves the way for new high-performance, end-to-end object detectors.