{"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/decomposing-motion-and-content-for-natural","title":"Decomposing Motion and Content for Natural Video Sequence Prediction","arxiv_id":"1706.08033","date":"2017-06-25","proceeding":null,"authors":["Ruben Villegas","Jimei Yang","Seunghoon Hong","Xunyu Lin","Honglak Lee"],"abstract":"We propose a deep neural network for the prediction of future frames in\nnatural video sequences. To effectively handle complex evolution of pixels in\nvideos, we propose to decompose the motion and content, two key components\ngenerating dynamics in videos. Our model is built upon the Encoder-Decoder\nConvolutional Neural Network and Convolutional LSTM for pixel-level prediction,\nwhich independently capture the spatial layout of an image and the\ncorresponding temporal dynamics. By independently modeling motion and content,\npredicting the next frame reduces to converting the extracted content features\ninto the next frame content by the identified motion features, which simplifies\nthe task of prediction. Our model is end-to-end trainable over multiple time\nsteps, and naturally learns to decompose motion and content without separate\ntraining. We evaluate the proposed network architecture on human activity\nvideos using KTH, Weizmann action, and UCF-101 datasets. We show\nstate-of-the-art performance in comparison to recent approaches. To the best of\nour knowledge, this is the first end-to-end trainable network architecture with\nmotion and content separation to model the spatiotemporal dynamics for\npixel-level future prediction in natural videos.","url_abs":"http://arxiv.org/abs/1706.08033v2","url_pdf":"http://arxiv.org/pdf/1706.08033v2.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":"decomposing-motion-and-content-for-natural","repo_url":"https://github.com/rubenvillegas/iclr2017mcnet","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"tf","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"future-prediction","task_name":"Future prediction"},{"task_slug":"prediction","task_name":"Prediction"},{"task_slug":"video-prediction","task_name":"Video Prediction"}],"methods":[{"method_slug":"lstm","method_name":"LSTM"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"tanh-activation","method_name":"Tanh Activation"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/video-prediction-on-kth","task":"Video Prediction","dataset":"KTH","model":"MCnet + Residual","rank_in_archive_order":24,"of":31,"metrics":{"Cond":"10","PSNR":"26.29","Pred":"20","SSIM":"0.806"},"uses_additional_data":false},{"leaderboard":"/sota/video-prediction-on-kth","task":"Video Prediction","dataset":"KTH","model":"MCnet","rank_in_archive_order":25,"of":31,"metrics":{"Cond":"10","PSNR":"25.95","Pred":"20","SSIM":"0.804"},"uses_additional_data":false}],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1706.08033","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}