Mutual attention inception network developed for remote sensing visual question answering

A probe squad led by Prof. Lu Xiaoqiang from the Xi'an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Sciences projected a caller communal attraction inception web (MAIN) and a dataset named RSIVQA for distant sensing ocular question answering. The results were published successful IEEE Transactions connected Geoscience and Remote Sensing. 

Remote sensing ocular question answering (VQA) chiefly aims astatine making semantic knowing of remote sensing images (RSIs) nonsubjective and interactive. Specifically, fixed an RSI, an intelligent cause volition reply a question astir the distant sensing scene.

Most of the existing methods disregard the spatial accusation of RSIs and word-level semantic accusation of questions, which restricts their applications successful galore analyzable scenes.

Accordingly, successful this study, the projected MAIN was made up of 2 parts, including the practice module and the fusion module. The practice module was devised to get the features of representation and question which tin supply amended representations.

As for the fusion module, it enhanced the discriminative quality of representations which tin get correct answers by reinforcing the representations of representation and question.

According to the experiments results, the projected method tin seizure the alignments betwixt images and questions nether antithetic valuation metrics. This survey provides a caller position for the distant sensing ocular question answering.

---

---

More information: Xiangtao Zheng et al, Mutual Attention Inception Network for Remote Sensing Visual Question Answering, IEEE Transactions connected Geoscience and Remote Sensing (2021). DOI: 10.1109/TGRS.2021.3079918