An anchor-free object detector based on soften optimized bi-directional FPN

Abstract

We propose an anchor-free object detector that combines a weighted bi-directional Feature Pyramid Network (BiFPN) and Soft Anchor Point Detector to address the object detection problem in a pixel-wise paradigm. The current mainstream object detection methods are anchor-based, which require to set hyper parameters such as scale and aspect ratio. This requires strong prior knowledge and can be difficult to design. Therefore, we propose an anchor-free detector that completely avoids the complex calculations and all the hyper parameters related to the anchor box by eliminating the predefined set of anchor boxes in an anchor-free way. Anchor-free detectors are essentially dense prediction methods. Although the huge solution space can yield high recall, simple anchor-free methods tend to return too many false positives, which leads to the problem of semantic ambiguity caused by the high overlap of object centers. Therefore, we propose BiFPN to alleviate the impact of high overlap which also effectively addresses the problems related to multi-scale features. Moreover, in order to utilize the power of feature pyramid better, we tackle the issues with a novel training strategy that involves two soften optimization techniques, i.e., soft-weighted anchor points and soft-selected pyramid levels. This training strategy further re-weights the quality of the detection results to make our detection results more stable.