Diffusion-based Data Generation for Out-of-Distribution Object Detection
Keywords: OOD detection, diffusion model, object detection
Abstract: Generating out-of-distribution (OOD) data is critical for training OOD object detectors, enabling them to identify OOD objects or categories as ``unknown''. Previous methods may generate imprecise OOD features due to incorrect assumptions on in-distribution (ID) data distribution. In this paper, we propose to discard any distribution assumption, leveraging a diffusion model to faithfully model the ID data distribution, and design a filtering strategy to generate accurate OOD data samples for training an unknown-aware object detector. Unlike previous methods that rely on predefined parametric models for modeling distributions, our diffusion model captures the latent feature distributions of ID data, which allows us to synthesize data samples within a compact feature space. We further design a filtering strategy based on K-Nearest Neighbors (KNN) to select low-density data samples proximate to the ID data as generated OOD samples, which are more challenging and effective for improving the OOD detector. Our method is generic and can be easily integrated with existing baseline methods, demonstrating superior performance on multiple benchmark datasets. The code will be made publicly available.
Supplementary Material: pdf
Primary Area: representation learning for computer vision, audio, language, and other modalities
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: 2812
Loading