{"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/show-adapt-and-tell-adversarial-training-of","title":"Show, Adapt and Tell: Adversarial Training of Cross-domain Image Captioner","arxiv_id":"1705.00930","date":"2017-05-02","proceeding":"ICCV 2017 10","authors":["Tseng-Hung Chen","Yuan-Hong Liao","Ching-Yao Chuang","Wan-Ting Hsu","Jianlong Fu","Min Sun"],"abstract":"Impressive image captioning results are achieved in domains with plenty of\ntraining image and sentence pairs (e.g., MSCOCO). However, transferring to a\ntarget domain with significant domain shifts but no paired training data\n(referred to as cross-domain image captioning) remains largely unexplored. We\npropose a novel adversarial training procedure to leverage unpaired data in the\ntarget domain. Two critic networks are introduced to guide the captioner,\nnamely domain critic and multi-modal critic. The domain critic assesses whether\nthe generated sentences are indistinguishable from sentences in the target\ndomain. The multi-modal critic assesses whether an image and its generated\nsentence are a valid pair. During training, the critics and captioner act as\nadversaries -- captioner aims to generate indistinguishable sentences, whereas\ncritics aim at distinguishing them. The assessment improves the captioner\nthrough policy gradient updates. During inference, we further propose a novel\ncritic-based planning method to select high-quality sentences without\nadditional supervision (e.g., tags). To evaluate, we use MSCOCO as the source\ndomain and four other datasets (CUB-200-2011, Oxford-102, TGIF, and Flickr30k)\nas the target domains. Our method consistently performs well on all datasets.\nIn particular, on CUB-200-2011, we achieve 21.8% CIDEr-D improvement after\nadaptation. Utilizing critics during inference further gives another 4.5%\nboost.","url_abs":"http://arxiv.org/abs/1705.00930v2","url_pdf":"http://arxiv.org/pdf/1705.00930v2.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":"show-adapt-and-tell-adversarial-training-of","repo_url":"https://github.com/tsenghungchen/show-adapt-and-tell","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":null}],"tasks":[{"task_slug":"image-captioning","task_name":"Image Captioning"},{"task_slug":"sentence","task_name":"Sentence"},{"task_slug":null,"task_name":"valid"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1705.00930","atlas_url":"https://app.syntology.ai/?focus=1705.00930","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}