A low-cost low-energy approach to VQA on traffic signs problems

For VLSP 2025 MLQA-TSR, we propose a simple retrieval-based pipeline that requires no model training. Text and image features are extracted using Jina Embeddings v3, C-RADIOv2-B, and Owlv2, then stored in Qdrant for cosine similarity search. Retrieved examples directly provide legal terms for subtask 1 and serve as few-shot prompts for Llama 4 Maverick in subtask 2.

Our method achieved a top-5 ranking (F2 = 0.54) for retrieval and a top-1 ranking (accuracy = 0.86) for question answering.

Setup

Conda environment

sh scripts/conda_setup.sh
conda activate mlqa-tsr
sh scripts/pip_setup.sh

Docker services

docker compose for qdrant

name: qdrant
services:
    qdrant:
        ports:
            - 6333:6333
            - 6334:6334
        volumes:
            - ${PWD}/qdrant_storage:/qdrant/storage:z
        image: qdrant/qdrant
        restart: always

For vllm, it really depends on the GPU infrastructure.

See vllm and qdrant docs

Run

At notebooks folder,

Read detect_image_object.ipynb, extract_image_feature.ipynb firstly, then run it

Read index_qdrant.ipynb secondly, then run it

Read naive_vector_search.ipynb for subtask 1, then run it

Read vlm_answer.ipynb for subtask 2, then run it

The src folder contains reusable code such as processing data, inferring models, connecting databases.

Citation

@inproceedings{anh-etal-2025-low,
    title = "A low-cost low-energy approach to {VQA} on traffic signs problems",
    author = "Anh, Vu Dinh  and
      Tran, Khiem Vinh  and
      Ha, Tran Thi",
    editor = "Mai, Luong Chi  and
      Huyen, Nguyen Thi Minh  and
      Trang, Nguyen Thi Thu",
    booktitle = "Proceedings of the 11th International Workshop on Vietnamese Language and Speech Processing",
    month = oct,
    year = "2025",
    address = "Hanoi, Vietnam",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.vlsp-1.53/",
    pages = "451--458"
}