CVPR Daily - Wednesday

by Liang Zheng, Australian National University The vision datasets understanding is one of the earliest data-centric workshops in CVPR. In the community, many people are excited about large-language models, diffusion models etc, but more are paying attention to data, an essential component in the computer vision system. The workshop of vision datasets understanding was thus created in 2022 to further spark interest from people in this area and received overwhelmingly positive feedback and attendance from researchers. Examples of focused problems are quality of training data, difficulty of test data, among other various dataset properties. This year we have eight oral presentations and two keynote speakers. Besides, an important change in this 2023 edition is the inclusion of DataCV challenge, where the task is to predict model accuracy in various test environments without access to the test ground truths. Our two keynote speakers have done outstanding works in data-centric AI fields. Vijay Janapa Redid from Harvard University talked about the Data- centric AI workshop and challenge happened at NeurIPS 2021. They developed the DataPerf platform to benchmark key stages in the data- centric pipeline to improve data quality. These stages include data parsing, quality assessment, cleaning, etc. James Zou from Stanford University discussed his data-centric works in medical analysis applications as well as fundamental machine learning theories. His group studies data valuation where they proposed a few interesting metrics, to provide statistical insights into which data are beneficial or detrimental to model training. The two talks received overwhelmingly positive feedback, where audience said “CVPR should have more such talks in the data-centric area” . Apart from these two talks, Haiyu Wu from University of Notre Dame presented his research analysis on the ambiguous and inconsistent annotations of the CelebA dataset, after which audience lined up to ask 18 DAILY CVPR Workshop Wednesday Liang Zheng (right in the photo with Andrew Ng) is a Senior Lecturer in the School of Computing, Australian National University (ANU)