SITTO: Single-Image Textured Mesh Reconstruction through Test-Time Optimization
Keywords: Feed-forward Mesh Prediction, Multi-View Diffusion, Test-Time Optimization, Textured Mesh Generation, DMTet, Physically Based Rendering (PBR)
Abstract: Reconstruction of a 3D textured mesh from a single image has been a long-standing and challenging problem. To address this challenge, we aim to leverage existing feed-forward-based models designed for predicting shape (i.e., textureless mesh) from a single image. However, there are difficulties that have to be overcome. Firstly, methods that estimate shape using feed-forward approaches cannot always guarantee high-quality results. A test-time optimization technique with feedback loops specified to each target object instance is necessary to apply these methods practically. To tackle this, we unlock the recent advancements in multi-view diffusion models, showing impressive multi-view image generation performances. Nonetheless, there are challenges associated with utilizing diffusion models. Specifically, it is crucial to estimate the viewpoint of the given reference image (i.e., its elevation and azimuth angles) and sample relative viewpoints from the reference viewpoint. We solely employ neural mesh representation and texture optimization to optimize training efficiency in terms of time and memory complexity. SITTO tackles these challenges by introducing an automatic pipeline for monocular 3D textured mesh reconstruction with test-time optimization. Our method demonstrates impressive results in fine-grained geometry details and the generation of realistic texture appearances.
Supplementary Material: zip
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 154
Loading