Unbiased VQA via modal information interaction and question transformation

Dahe Peng, Zhixin Li

Abstract: Highlights•We propose a unbiased VQA method IEQM to solve the language prior problem.•We design two modules ETV and INV to generate two high-level visual features.•We design a question Converter that converts original question into three sub-questions.•We design a joint training method to jointly train IAE and QC modules.