Hierarchical Regression and Classification for Accurate Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Single-shot detectors (SSDs) commonly use fully convolutional networks for simultaneous object classification and bounding box regression.
• Existing SSD architectures suffer from classification mismatch and insufficient localization accuracy due to single-regression steps.
• These limitations hinder the performance of real-time object detection systems.

Purpose of the Study:

• To propose a novel reg-offset-cls (ROC) module to enhance object detection accuracy.
• To improve both classification and localization of objects in images.
• To achieve state-of-the-art performance in one-stage object detection at real-time speeds.

Main Methods:

• Introduced a novel reg-offset-cls (ROC) module with three hierarchical steps: bounding box regression, feature sampling, and classification.
• Stacked two ROC modules, with the second module taking the output of the first as input, for improved localization.
• Incorporated a feature enhanced (FE) module between the stacked ROC modules to extract richer contextual information.

Main Results:

• The proposed method demonstrated superior performance on MS COCO, PASCAL VOC, and UAVDT datasets.
• Outperformed existing state-of-the-art one-stage object detection methods.
• Achieved real-time detection speeds without additional complex components.

Conclusions:

• The novel ROC module effectively addresses classification mismatch and enhances localization accuracy in object detection.
• Stacking ROC modules with an integrated FE module significantly improves detection performance.
• The proposed method offers a highly effective and efficient solution for real-time object detection.