{"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/marrnet-3d-shape-reconstruction-via-25d","title":"MarrNet: 3D Shape Reconstruction via 2.5D Sketches","arxiv_id":"1711.03129","date":"2017-11-08","proceeding":"NeurIPS 2017 12","authors":["Jiajun Wu","Yifan Wang","Tianfan Xue","Xingyuan Sun","William T. Freeman","Joshua B. Tenenbaum"],"abstract":"3D object reconstruction from a single image is a highly under-determined\nproblem, requiring strong prior knowledge of plausible 3D shapes. This\nintroduces challenges for learning-based approaches, as 3D object annotations\nare scarce in real images. Previous work chose to train on synthetic data with\nground truth 3D information, but suffered from domain adaptation when tested on\nreal data. In this work, we propose MarrNet, an end-to-end trainable model that\nsequentially estimates 2.5D sketches and 3D object shape. Our disentangled,\ntwo-step formulation has three advantages. First, compared to full 3D shape,\n2.5D sketches are much easier to be recovered from a 2D image; models that\nrecover 2.5D sketches are also more likely to transfer from synthetic to real\ndata. Second, for 3D reconstruction from 2.5D sketches, systems can learn\npurely from synthetic data. This is because we can easily render realistic 2.5D\nsketches without modeling object appearance variations in real images,\nincluding lighting, texture, etc. This further relieves the domain adaptation\nproblem. Third, we derive differentiable projective functions from 3D shape to\n2.5D sketches; the framework is therefore end-to-end trainable on real images,\nrequiring no human annotations. Our model achieves state-of-the-art performance\non 3D shape reconstruction.","url_abs":"http://arxiv.org/abs/1711.03129v1","url_pdf":"http://arxiv.org/pdf/1711.03129v1.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":[],"tasks":[{"task_slug":"3d-object-reconstruction","task_name":"3D Object Reconstruction"},{"task_slug":"3d-object-reconstruction-from-a-single-image","task_name":"3D Object Reconstruction From A Single Image"},{"task_slug":"3d-reconstruction","task_name":"3D Reconstruction"},{"task_slug":"3d-shape-reconstruction","task_name":"3D Shape Reconstruction"},{"task_slug":"domain-adaptation","task_name":"Domain Adaptation"},{"task_slug":"object","task_name":"Object"},{"task_slug":"object-reconstruction","task_name":"Object Reconstruction"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/3d-shape-retrieval-on-pix3d","task":"3D Shape Classification","dataset":"Pix3D","model":"MarrNet","rank_in_archive_order":2,"of":3,"metrics":{"R@1":"0.42","R@16":"0.71","R@2":"0.51","R@32":"0.78","R@4":"0.57","R@8":"0.64"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1711.03129","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}