{"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/ti-pooling-transformation-invariant-pooling","title":"TI-POOLING: transformation-invariant pooling for feature learning in Convolutional Neural Networks","arxiv_id":"1604.06318","date":"2016-04-21","proceeding":"CVPR 2016 6","authors":["Dmitry Laptev","Nikolay Savinov","Joachim M. Buhmann","Marc Pollefeys"],"abstract":"In this paper we present a deep neural network topology that incorporates a\nsimple to implement transformation invariant pooling operator (TI-POOLING).\nThis operator is able to efficiently handle prior knowledge on nuisance\nvariations in the data, such as rotation or scale changes. Most current methods\nusually make use of dataset augmentation to address this issue, but this\nrequires larger number of model parameters and more training data, and results\nin significantly increased training time and larger chance of under- or\noverfitting. The main reason for these drawbacks is that the learned model\nneeds to capture adequate features for all the possible transformations of the\ninput. On the other hand, we formulate features in convolutional neural\nnetworks to be transformation-invariant. We achieve that using parallel siamese\narchitectures for the considered transformation set and applying the TI-POOLING\noperator on their outputs before the fully-connected layers. We show that this\ntopology internally finds the most optimal \"canonical\" instance of the input\nimage for training and therefore limits the redundancy in learned features.\nThis more efficient use of training data results in better performance on\npopular benchmark datasets with smaller number of parameters when comparing to\nstandard convolutional neural networks with dataset augmentation and to other\nbaselines.","url_abs":"http://arxiv.org/abs/1604.06318v2","url_pdf":"http://arxiv.org/pdf/1604.06318v2.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":"ti-pooling-transformation-invariant-pooling","repo_url":"https://github.com/nsavinov/semantic3dnet","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"torch","reach":{"status":"unanswered"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1604.06318","atlas_url":"https://app.syntology.ai/?focus=1604.06318","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}