Afmf: Adaptive Fusion of Multi-Scale Features for Pixel-Level Object Detection

Abstract

Small targets have the characteristics of low resolution and low amounts of feature information, leading to the weak expression ability of extracted features, which greatly hinders object detection accuracy. This paper adopts the pixel-level prediction and regression method based on fully convolutional networks (FCNs) to establish the one-stage anchor-free object detection model named AFMF. The Adaptive Spatial Pyramid Pooling (ASPP) module and Adaptive Spatial Fusion Pyramid (ASFP) module are proposed for this model. The ASPP module, which is attached to the backbone network, can obtain more fine-grained features by enlarging the receptive field of the original features and adaptively aggregating the features of different receptive fields to enrich the context information of local areas. The adaptive weighted fusion in the ASFP module was applied to the multiscale features to obtain the feature pyramid with more semantic information. Meanwhile, a residual connection is added to obtain spatial context information with ratio-invariance to reduce the loss of location information of the original features. In the single-model and single-scale test, the AFMF detector uses ResNext-64x4d-101 to achieve 44.3% AP on the MS COCO dataset, surpassing the previous anchor-free one-stage detector based on FCN and maintaining real-time detection.

Keywords: Adaptive Spatial Pyramid Pooling, Adaptive Feature Pyramid Network, Multi-scale context information, Residual connection