LGADet: Light-weight Anchor-free Multispectral Pedestrian Detection with Mixed Local and Global Attention
Lists
Only you see this
Abstract
Balancing accuracy and efficiency is of significant importance for multispectral pedestrian detection in practical applications. To address these problems, a light-weight anchor-free multispectral pedestrian detection method with mixed Local and Global Attention mechanism (LGA) is proposed to narrow the gap between academic research and practical application. The anchor-free detection pipeline equipped with light-weight backbone leads to significant speedup, while a mixed attention mechanism is utilized to refine features in order to improve the accuracy. Specifically, an anchor-free pedestrian detection framework with MobileNetV2 backbone is firstly utilized to reduce the computational complexity, achieving significant speedup for model inference. Secondly, our method makes use of DMAF module to enhance complementary information between RGB and Thermal image features. Finally, the quality of feature fusion is greatly improved with local and global attention mechanisms, thus enhancing the detection accuracy. Experiments on the KAIST, FLIR and CVC-14 datasets show significant performance improvement in terms of MR, comparing with other state-of-the-art methods. When deployed on the Nvidia Jetson TX2, impressing result is obtained with good compromise between accuracy and speed.
BibTeX
@inproceedings{zuo2023lgadet,
author = {Xin Zuo and Zhi Wang and Yue Liu and Jifeng Shen and Haoran Wang},
title = {{LGADet: Light-weight Anchor-free Multispectral Pedestrian Detection with Mixed Local and Global Attention}},
booktitle = {Neural Process. Lett},
year = {2023},
url = {https://openreview.net/forum?id=IswQ9SEcjd},
}Loading related content...