Object Detection with OOD Generalizable Neural Architecture Search

Fan Wu, Nanyang Ye, Lanqing HONG, Zhenguo Li, Chensheng Peng

Keywords: Out-of-Distribution, neural architecture search

Abstract: We present a Neural Architecture Search (NAS) framework guided by feature orthogonalization to improve Out-of-Distribution (OOD) Generalization on Object Detection. Specifically, we attribute the failure of generalizing on OOD data to the spurious correlations of category-related features and context-related features. The category-related features describe the causal information for predicting the target objects, e.g., "a car with four wheels", while the context-related features describe the non-causal information, e.g., "a car driving at night", and the context-related features are always mistaken for causal information due to the existence of distinct data distribution between training and testing sets (OOD) to some degree. Therefore, we aim at automatically discovering an optimal architecture that is able to disentangle the category-related features and the context-related features with a novel weight-based detector head. Both theoretical and experimental results show that the proposed scheme is able to achieve the disentanglement and better performance on both Independent-Identically-Distribution datasets (Pascal VOC 2012 and MS COCO) and OOD datasets (BDD100K-weather and BDD100K-time-of-day).

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)