Abstract
Visual question answering (VQA) has been attracting attention in remote sensing very recently. However, the proposed solutions remain rather limited in the sense that the existing VQA datasets address closed-ended question-answer queries, which may not necessarily reflect real open-ended scenarios. In this paper, we propose a new dataset named VQA-TextRS that was built manually with human annotations and considers various forms of open-ended question-answer pairs. Moreover, we propose an encoder-decoder architecture via transformers on account of their self-attention property that allows relational learning of different positions of the same sequence without the need of typical recurrence operations. Thus, we employed vision and natural language processing (NLP) transformers respectively to draw visual and textual cues from the image and respective question. Afterwards, we applied a transformer decoder, which enables the cross-attention mechanism to fuse the earlier two modalities. The fusion vectors correlate with the process of answer generation to produce the final form of the output. We demonstrate that plausible results can be obtained in open-ended VQA. For instance, the proposed architecture scores an accuracy of 84.01% on questions related to the presence of objects in the query images.