eprintid: 10074121 rev_number: 18 eprint_status: archive userid: 608 dir: disk0/10/07/41/21 datestamp: 2019-05-16 16:26:52 lastmod: 2021-11-15 01:55:19 status_changed: 2019-05-16 16:26:52 type: proceedings_section metadata_visibility: show creators_name: Firman, M creators_name: Campbell, NDF creators_name: Agapito, L creators_name: Brostow, GJ title: DiverseNet: When One Right Answer is not Enough ispublished: pub divisions: UCL divisions: B04 divisions: C05 divisions: F48 keywords: Training, Task analysis, Aerospace electronics, Three-dimensional displays, Supervised learning, Two dimensional displays, Training data note: This version is the author accepted manuscript. For information on re-use, please refer to the publisher’s terms and conditions. abstract: Many structured prediction tasks in machine vision have a collection of acceptable answers, instead of one definitive ground truth answer. Segmentation of images, for example, is subject to human labeling bias. Similarly, there are multiple possible pixel values that could plausibly complete occluded image regions. State-of-the art supervised learning methods are typically optimized to make a single test-time prediction for each query, failing to find other modes in the output space. Existing methods that allow for sampling often sacrifice speed or accuracy. We introduce a simple method for training a neural network, which enables diverse structured predictions to be made for each test-time query. For a single input, we learn to predict a range of possible answers. We compare favorably to methods that seek diversity through an ensemble of networks. Such stochastic multiple choice learning faces mode collapse, where one or more ensemble members fail to receive any training signal. Our best performing solution can be deployed for various tasks, and just involves small modifications to the existing single-mode architecture, loss function, and training regime. We demonstrate that our method results in quantitative improvements across three challenging tasks: 2D image completion, 3D volume estimation, and flow prediction. date: 2018-12-17 date_type: published publisher: IEEE official_url: https://doi.org/10.1109/CVPR.2018.00587 oa_status: green full_text_type: other language: eng primo: open primo_central: open_green verified: verified_manual elements_id: 1633661 doi: 10.1109/CVPR.2018.00587 isbn_13: 978-1-5386-6420-9 lyricists_name: Brostow, Gabriel lyricists_name: De Agapito Vicente, Lourdes lyricists_id: GBROS38 lyricists_id: LDEAG40 actors_name: Brostow, Gabriel actors_id: GBROS38 actors_role: owner full_text_status: public series: IEEE/CVF Conference on Computer Vision and Pattern Recognition publication: 2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) volume: 2018 place_of_pub: Salt Late City, UT, USA pagerange: 5598-5607 pages: 10 event_title: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition event_location: Salt Lake City, UT event_dates: 18 June 2018 - 23 June 2018 institution: 31st IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) issn: 1063-6919 book_title: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition citation: Firman, M; Campbell, NDF; Agapito, L; Brostow, GJ; (2018) DiverseNet: When One Right Answer is not Enough. In: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. (pp. pp. 5598-5607). IEEE: Salt Late City, UT, USA. Green open access document_url: https://discovery.ucl.ac.uk/id/eprint/10074121/1/cvpr18_diversenet.pdf