Enhancing Geometric Factors in Model Learning and Inference for Object Detection and Instance Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep learning models have significantly advanced object detection and instance segmentation.
• Current methods often struggle with accurately capturing geometric relationships crucial for precise localization.

Purpose of the Study:

• To enhance bounding-box regression and non-maximum suppression (NMS) by incorporating key geometric factors.
• To improve object detection and instance segmentation performance without compromising inference efficiency.

Main Methods:

• Proposed Complete-IoU (CIoU) loss, integrating overlap area, normalized central-point distance, and aspect ratio for improved bounding-box regression.
• Introduced Cluster-NMS, an efficient GPU-implemented NMS technique that implicitly clusters detected boxes, reducing iterations and enhancing geometric factor incorporation.

Main Results:

• CIoU loss consistently improved average precision (AP) and average recall (AR) compared to existing losses.
• Cluster-NMS demonstrated efficiency and effectiveness in improving both AP and AR.
• Applied to state-of-the-art models like YOLACT, YOLOv3, SSD, and Faster R-CNN, achieving notable performance gains (e.g., +1.7 AP, +6.2 AR100 for object detection on YOLACT).

Conclusions:

• CIoU loss and Cluster-NMS offer significant improvements in object detection and instance segmentation by effectively utilizing geometric information.
• The proposed methods provide a robust and efficient approach to enhance deep learning models for visual recognition tasks.