MRNet: Multiple-Input Receptive Field Network for Large-Scale Point Cloud Segmentation

The size of the input receptive field is one of the most critical aspects in the semantic segmentation of the point cloud, yet it is one of the most overlooked parameters. This paper presents the multiple-input receptive field processing semantic segmentation network MRNet. The fundamental philosophy of our design is to overcome the size of the input receptive field dilemma. In particular, the input receptive field's size significantly impacts the performance of different sizes of objects. To overcome this, we introduce a parallel processing network with connection modules between the parallel streams. Our ablation studies show the effectiveness of implemented modules. Also, we set the new state-of-art performance on the large-scale point cloud dataset SensatUrban.