BMFN3D: Bidirectional multilayer fusion network for indoor 3D object detection

Abstract

Among various indoor 3D object detection methods, one of which is based on 3D convolution. However, in current 3D convolutional method for indoor 3D object detection, feature fusion is only implemented in a single direction and adjacent layer, which leads to the feature loss during feature propagation and weakens the detection performance. To solve the above problem, a bidirectional multilayer fusion network (BMFN) is proposed to enrich features for prediction. Furthermore, we propose a new network based on BMFN for 3D object detection, BMFN3D. In BMFN3D, the channel and spatial attention mechanisms are introduced into the backbone and head of detection network, so as to enhance the selection of effective feature information. Finally, a new calculation method for regression loss is proposed with the consideration of the 3D spatial information. We validate our method with two indoor datasets of SUN RGB-D and ScanNet, and achieve state-of-the-art results.