method: BERT Large Ensemble2023-03-18

Authors: ZhuangZhuang Cai

Affiliation: GammaLab, Pingan


Description: News video question-answering method based on OCR, layout analysis, object tracking, and ASR technologies. The method utilizes OCR technology to recognize text in video frames, uses layout analysis to merge paragraphs, employs object tracking algorithms to remove duplicate text in video frames, and finally uses ASR technology to transcribe speech in video clips. The OCR de-duplicated text and ASR text are concatenated to form the context for an extractive question-answering task,we then fine-tuned the pre-trained model. Our method achieved competitive results in the ICDAR2023 NewsVideoQA competition, demonstrating the effectiveness of using OCR and ASR technologies for news video question-answering.

Ranking Table

Description Paper Source Code
2023-03-18BERT Large Ensemble0.72260.6251

Ranking Graphic

Ranking Graphic