Top1 Solution of QQ Browser 2021 Ai Algorithm Competition Track 1 : Multimodal Video Similarity

TL;DR

This paper presents the winning solution to a video similarity competition using a multi-modal transformer trained with multiple tasks, achieving top leaderboard performance and releasing source code.

Contribution

It introduces a multi-modal transformer approach with multi-task pretraining and fine-tuning for video similarity, winning a major AI competition.

Findings

Achieved 0.852 score on leaderboard

Ensembled multiple models for top performance

Source code released for reproducibility

Abstract

In this paper, we describe the solution to the QQ Browser 2021 Ai Algorithm Competition (AIAC) Track 1. We use the multi-modal transformer model for the video embedding extraction. In the pretrain phase, we train the model with three tasks, (1) Video Tag Classification (VTC), (2) Mask Language Modeling (MLM) and (3) Mask Frame Modeling (MFM). In the finetune phase, we train the model with video similarity based on rank normalized human labels. Our full pipeline, after ensembling several models, scores 0.852 on the leaderboard, which we achieved the 1st place in the competition. The source codes have been released at Github.