CIDA3D: Conformal Inference aided unsupervised Domain Adaptation for 3D-Aware Classification
Keywords: UDA, 3D pose estimation, 3D-Aware classification, occlusion, robustness
TL;DR: Image Only UDA for 3D Aware Classification and MultiClass 3D Pose Estimation
Abstract: Cognitive Science studies show that human perception becomes robust to occlusions and other nuisances due to internal 3D representations of objects. This idea has been incorporated into computer vision models to improve their ability to understand and reason about the 3D world. However, collecting 3D annotations in vision datasets is expensive. This makes the robustness of the perception model to distribution shifts challenging. We introduce Conformal Inference aided unsupervised Domain Adaptation (CIDA)-3D for the complex setting of multiclass pose estimation. Our method adapts category level pose estimation (3D) models in nuisance ridden target domains directly from images without class label information, by harnessing uncertainty in model predictions (using conformal sets). This allows for significantly better and computationally efficient adaptation to target domains with synthetic and real-world noise. We also show a robust adaptation from fully synthetic data to complex real-world domains. To the best of our knowledge, this method is the first to attempt unsupervised domain adaptation for robust 3D-aware classification and multiclass pose estimation in real-world scenarios by adapting models trained on procedurally generated synthetic data.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 4571
Loading