A Multi-Level Semantic Fusion VoteNet for 3D Object Detection on Point Clouds

In this paper, a Multi-Level Semantic Fusion VoteNet (MLSFVNet) is proposed to detect objects in 3D scenes. The method works on 3D point clouds captured by RGBD camera, which can provide abundant and precise distance information of environments. The proposed method consists of three modules: the multi-level semantics fusion network, voting operation and proposal generator. To overcome the lack of semantic information, the multi-level semantics fusion network is proposed to capture the multi-level features. To predict the object centers, the voting operation is used to map the features into a feature space of the same scale and regress the object centers. The proposal generator is used to generate proposals and then predict the bounding boxes. MLSFVNet is evaluated on the popular indoor datasets SUN RGB-D and ScanNetV2. The experimental results demonstrate that the MLSFVNet proposed in this paper is an effective way to promote detection accuracy: 58.1% mAP on SUN RGB-D and 59.8% mAP on ScanNetV2.