{"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/recombinator-networks-learning-coarse-to-fine","title":"Recombinator Networks: Learning Coarse-to-Fine Feature Aggregation","arxiv_id":"1511.07356","date":"2015-11-23","proceeding":"CVPR 2016 6","authors":["Sina Honari","Jason Yosinski","Pascal Vincent","Christopher Pal"],"abstract":"Deep neural networks with alternating convolutional, max-pooling and\ndecimation layers are widely used in state of the art architectures for\ncomputer vision. Max-pooling purposefully discards precise spatial information\nin order to create features that are more robust, and typically organized as\nlower resolution spatial feature maps. On some tasks, such as whole-image\nclassification, max-pooling derived features are well suited; however, for\ntasks requiring precise localization, such as pixel level prediction and\nsegmentation, max-pooling destroys exactly the information required to perform\nwell. Precise localization may be preserved by shallow convnets without pooling\nbut at the expense of robustness. Can we have our max-pooled multi-layered cake\nand eat it too? Several papers have proposed summation and concatenation based\nmethods for combining upsampled coarse, abstract features with finer features\nto produce robust pixel level predictions. Here we introduce another model ---\ndubbed Recombinator Networks --- where coarse features inform finer features\nearly in their formation such that finer features can make use of several\nlayers of computation in deciding how to use coarse features. The model is\ntrained once, end-to-end and performs better than summation-based\narchitectures, reducing the error from the previous state of the art on two\nfacial keypoint datasets, AFW and AFLW, by 30\\% and beating the current\nstate-of-the-art on 300W without using extra data. We improve performance even\nfurther by adding a denoising prediction model based on a novel convnet\nformulation.","url_abs":"http://arxiv.org/abs/1511.07356v2","url_pdf":"http://arxiv.org/pdf/1511.07356v2.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":[{"paper_slug":"recombinator-networks-learning-coarse-to-fine","repo_url":"https://github.com/SinaHonari/RCN","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok","spdx":"NOASSERTION"}}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"image-classification","task_name":"image-classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1511.07356","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}