2019/06/26 by Davy Neven, Neven, Davy, Bert De Brabandere +5
Computer Science · Environmental Science · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Remote Sensing and LiDAR Applications
paper · pdf · doi:10.48550/arxiv.1906.11109
openalex publication_date 2019/06/26 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Current state-of-the-art instance segmentation methods are not suited for\nreal-time applications like autonomous driving, which require fast execution\ntimes at high accuracy. Although the currently dominant proposal-based methods\nhave high accuracy, they are slow and generate masks at a fixed and low\nresolution. Proposal-free methods, by contrast, can generate masks at high\nresolution and are often faster, but fail to reach the same accuracy as the\nproposal-based methods. In this work we propose a new clustering loss function\nfor proposal-free instance segmentation. The loss function pulls the spatial\nembeddings of pixels belonging to the same instance together and jointly learns\nan instance-specific clustering bandwidth, maximizing the\nintersection-over-union of the resulting instance mask. When combined with a\nfast architecture, the network can perform instance segmentation in real-time\nwhile maintaining a high accuracy. We evaluate our method on the challenging\nCityscapes benchmark and achieve top results (5 % improvement over Mask R-CNN)\nat more than 10 fps on 2MP images. Code will be available at\nhttps://github.com/davyneven/SpatialEmbeddings .\n