O ut O f D istribution Generalization in C omputer V ision

in conjunction with ICCV 2023

Deep learning models are usually developed and tested under the implicit assumption that the training and test data are drawn independently and identically distributed (IID) from the same distribution. Overlooking out-of-distribution (OOD) images can result in poor performance in unseen or adverse viewing conditions, which is common in real-world scenarios.

In this workshop, we are interested in discussing the performance of computer vision models on OOD images which follows a different distribution than the training images. Also given the current trend of web-scale pretrained computer vision models, it is of interest to better understand their performance in the OOD or rare scenarios.

Our workshop will be featuring three competitions, OOD generalization on the OOD-CV dataset, open-set recognition and generalized category discovery on the Semantic-shift benchmark.

Submission Information

We invite submissions of short papers on the topic of out-of-distribution generalization in computer vision. The short papers should use the CVPR template and be max. 4 pages long. The topics include but are not limited to:

Discussion of OOD generalization in the context of internet scale pretrained models
Improving generalization of computer vision systems in OOD scenarios
Research at the intersection of biological and machine vision
Generative causal models for image analysis
Domain generalization
Novel architectures with robustness to occlusion, viewpoint and other real-world domain shifts
Domain adaptation techniques for robust vision system in the real world
Datasets for evaluating model robustness

Please see here for submission informations.

Challenge Information

This workshop will feature three challenges:

Please see here for more information.

Each challenge will have two tracks, one track will limit the use of pretrained models to self-supervised pretrained models on ImageNet-1k and the other track will allow the use of any self-supervised pretrained models.

The top-3 teams are required to open source their code and models after the competition to ensure reproducibility.