Diverse Explanations for Object Detectors with Nesterov-Accelerated iGOS++
Mingqi Jiang (Oregon State University),* Saeed Khorram (Oregon State University), Li Fuxin (Oregon State University)
The 34th British Machine Vision Conference
Abstract
Object detection is a crucial task in computer vision, with applications ranging from autonomous driving to surveillance systems. However, few have approached the problem of explaining object detections to gain more insights. In this paper, we extend iGOS++, an explanation algorithm of image classification models, to the task of object detection. Our extension consists of two novel aspects. The first is to utilize Nesterov Accelerated Gradient (NAG) to improve the optimization with integrated gradients. This significantly improves over the line search used in the original work in terms of both speed and quality. Besides, we propose to generate diverse explanations via different initializations of the optimization algorithm, which can better showcase the robustness of the network under different occlusions. To evaluate the effectiveness of our algorithm, we conduct experiments on the MS COCO and PASCAL VOC datasets. Results demonstrate that our approach significantly outperforms existing methods in terms of both explanation quality and speed. Besides, the diverse explanations it generates give more insight into the (sometimes erroneous) mechanisms underlying deep object detectors.Video
Citation
@inproceedings{Jiang_2023_BMVC,
author = {Mingqi Jiang and Saeed Khorram and Li Fuxin},
title = {Diverse Explanations for Object Detectors with Nesterov-Accelerated iGOS++},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0188.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by