Object detection based on multiple trick feature pyramid networks and dynamic balanced L1 loss

Abstract

Although the performance of the object detection has been significantly optimised in recent years, there is still a lot of room for designing multi-scale feature fusion methods and designing loss functions. Specifically, we propose Multiple Trick Feature Pyramid Networks (MT-FPN), by using various techniques such as feedback information, global module, attention mechanism, and fusion of refined information, to solve the problem of insufficient multi-scale feature fusion. We also propose Dynamic Balanced L1 Loss (DBLL), by utilising dynamic strategies and solving the derivative discontinuity problem, in order to help relieve the inconsistent problem between the dynamic training process and the fixed parameters. Moreover, by replacing FPN with MT-FPN, our Average Precision (AP) on Microsoft Common Objects in Context (MSCOCO) is 5.1 points and 3.8 points higher than FPN Faster R-CNN and Libra R-CNN, respectively. Without any bells and whistles, our experiments also show that the combined application of MT-FPN and DBLL achieves competitive performance compared with most advanced detectors on MS COCO benchmark.