Feature Pyramid Reconfiguration with Consistent Loss for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection performance relies heavily on feature pyramids, but current methods struggle with integrating semantic information across scales.
• Existing object detectors face challenges with accurate localization due to coarse definitions of positive examples during training and prediction.

Purpose of the Study:

• To propose a novel architecture for object detection that effectively integrates semantic information across feature scales.
• To address the inaccurate object localization problem by refining the training loss function.

Main Methods:

• Developed a novel architecture featuring two lightweight, trainable processes: global attention and local reconfiguration, to enhance feature hierarchy.
• Proposed a modified cross-entropy loss function to improve the focus on accurate object predictions during training.
• Integrated these methods into popular one-stage (SSD, RetinaNet) and two-stage (Faster R-CNN) object detection frameworks.

Main Results:

• The proposed architecture significantly boosts object detection performance by effectively integrating multi-scale semantic information.
• The refined loss function demonstrably improves the accuracy of object localization.
• Experiments on PASCAL VOC 2007, PASCAL VOC 2012, and MS COCO datasets show consistent and significant improvements over state-of-the-art methods.

Conclusions:

• The novel feature reconfiguration and consistent loss function offer a flexible and effective approach to enhance object detection systems.
• The proposed methods provide substantial performance gains and improved localization accuracy, applicable to various existing detection frameworks.