Tracked-Vehicle Retrieval by Natural Language Descriptions with Multi-Contextual Adaptive Knowledge

This paper introduces our solution for Track 2 in AI City Challenge 2023. The task is tracked-vehicle retrieval by natural language descriptions with a real-world dataset of various scenarios and cameras. Our solution mainly focuses on four points: (1) To address the linguistic ambiguity in the language query, we leverage our proposed standardized version for text descriptions for the domain-adaptive training and post-processing stage. (2) Our baseline vehicle retrieval model utilizes CLIP to extract robust visual and textual feature representations to learn the unified cross-modal representations between textual and visual features. (3) Our proposed semi-supervised domain adaptive (SSDA) training method is leveraged to address the domain gap between the train and test set. (4) Finally, we propose a multi-contextual post-processing technique that prunes out the wrong results based on multi-contextual attributes information that effectively boosts the final retrieval results. Our proposed framework has yielded a competitive performance of 82.63% MRR accuracy on the test set, achieving 1st place in the competition. Codes will be available at https://github.com/zef1611/AIC23_NLRetrieval_HCMIU_CVIP