IMFNet: Interpretable Multimodal Fusion for Point Cloud Registration

The existing state-of-the-art point descriptor relies on structure information only, which omits the texture information. However, is crucial for our humans to distinguish a scene part. Moreover, current learning-based descriptors are all black boxes unclear how original points contribute final descriptor. This paper proposes new multimodal fusion method generate cloud registration by considering and Specifically, novel attention-fusion module designed extract weighted extraction. In addition, we propose an interpretable explain neural network visually showing contributing We use descriptor's channel value as loss backpropagate target layer consider gradient significance of this moves one step further explainable deep learning in task. Comprehensive experiments 3DMatch, 3DLoMatch KITTI demonstrate that achieves accuracy improves distinctiveness. also explaining