{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/decoders-matter-for-semantic-segmentation","title":"Decoders Matter for Semantic Segmentation: Data-Dependent Decoding Enables Flexible Feature Aggregation","arxiv_id":"1903.02120","date":"2019-03-05","proceeding":"CVPR 2019 6","authors":["Zhi Tian","Tong He","Chunhua Shen","Youliang Yan"],"abstract":"Recent semantic segmentation methods exploit encoder-decoder architectures to\nproduce the desired pixel-wise segmentation prediction. The last layer of the\ndecoders is typically a bilinear upsampling procedure to recover the final\npixel-wise prediction. We empirically show that this oversimple and\ndata-independent bilinear upsampling may lead to sub-optimal results.\n  In this work, we propose a data-dependent upsampling (DUpsampling) to replace\nbilinear, which takes advantages of the redundancy in the label space of\nsemantic segmentation and is able to recover the pixel-wise prediction from\nlow-resolution outputs of CNNs. The main advantage of the new upsampling layer\nlies in that with a relatively lower-resolution feature map such as\n$\\frac{1}{16}$ or $\\frac{1}{32}$ of the input size, we can achieve even better\nsegmentation accuracy, significantly reducing computation complexity. This is\nmade possible by 1) the new upsampling layer's much improved reconstruction\ncapability; and more importantly 2) the DUpsampling based decoder's flexibility\nin leveraging almost arbitrary combinations of the CNN encoders' features.\nExperiments demonstrate that our proposed decoder outperforms the\nstate-of-the-art decoder, with only $\\sim$20\\% of computation. Finally, without\nany post-processing, the framework equipped with our proposed decoder achieves\nnew state-of-the-art performance on two datasets: 88.1\\% mIOU on PASCAL VOC\nwith 30\\% computation of the previously best model; and 52.5\\% mIOU on PASCAL\nContext.","url_abs":"http://arxiv.org/abs/1903.02120v3","url_pdf":"http://arxiv.org/pdf/1903.02120v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"segmentation","task_name":"Segmentation"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/semantic-segmentation-on-pascal-context","task":"Semantic Segmentation","dataset":"PASCAL Context","model":"DUpsampling","rank_in_archive_order":49,"of":66,"metrics":{"mIoU":"52.5"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/1903.02120","atlas_url":"https://app.syntology.ai/?focus=1903.02120","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}