SINGLE-IMAGE COHERENT RECONSTRUCTION OF OBJECTS AND HUMANS
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.
Keywords: Scene Reconstruction, Mesh Collisions, Human-Human Interactions, Human-Object Interactions, Image Inpainting, Pose Estimation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Existing methods for reconstruction of objects and humans from a monocular image suffer from severe mesh collisions and performance limitations for interacting occluding objects. In this paper, we introduce a method that deduces spatial configurations and achieves globally consistent 3D reconstruction for interacting objects and people captured within a single image. Our contributions encompass: 1) an optimization framework, featuring a novel collision loss, tailored to handle complex human-object and human-human interactions, ensuring spatially coherent scene reconstruction; and 2) a novel technique for robustly estimating 6 degrees of freedom (DOF) poses, particularly for heavily occluded objects, exploiting image inpainting. Notably, our proposed method operates effectively on images from real-world scenarios, without necessitating scene or object-level 3D supervision. Through both qualitative and quantitative assessments, we demonstrate the superior quality of our reconstructions, showcasing a significant reduction in collisions in scenes with multiple interacting humans and objects.
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: 5662
Loading