Efficient Object Detection Using Total Energy Function: An Alternative to Anchor Box

Bounding box regression is a popular technique for fine-tuning or predicting the localization boxes in object detection methods that are trained to regress from proposed regions or fixed anchor boxes to bounding boxes.But anchor-based methods face computational costs and challenges in optimal anchor configuration, which constrain their adaptability and performance.This paper proposes an anchor-free object detection method using multi-component energy optimization that integrates edge, region, size, and shape information to determine the optimal bounding box parameters.Experiments on PASCAL VOC 2012 show superior performance with 85.7% loss reduction for small objects and 3-4 times faster convergence speed, demonstrating the effectiveness of the energy optimization method.