Arxiv link: arxiv.org/abs/2203.15395
TLDR: We proposed a metric to evaluate societal bias in image captioning and found that all the models (including gender equalisers) amplify gender and racial bias.
Our paper Quantifying Societal Bias in Image Captioning has been accepted as an oral at #CVPR2022@CVPR - by Yusuke Hirota for his master course project, Yuta Nakashima, and myself at @ids_ou. Arxiv version coming up soon!
Glad to share our recent work on explainable AI.
SCOUTER🧐: Slot Attention-based Classifier for Explainable Image Recognition
SCOUTER performs classification by answering "why the image is of this category" or "why it is not."
PDF and code are available.
arxiv.org/abs/2009.06138
Our paper on knowledge-based VideoQA has been accepted at #ECCV2020. We generate unsupervised descriptions of video scenes as our video representation. Arxiv paper coming soon!