dlgpr

DQ-Pilot (DlgPR)

This repository contains the final merged SFT+GRPO DQ-Pilot checkpoint from DialogueVPR: Towards Conversational Visual Place Recognition. DQ-Pilot examines candidate street-view images and the dialogue history, then asks a discriminative follow-up question to improve place retrieval.

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "graysonggg/dlgpr"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

The complete CMPL retrieval and dialogue pipeline is available in the official code repository. Training artifacts are available in the associated Hugging Face data repository.

Citation

@inproceedings{song2026dialoguevpr,
  title={DialogueVPR: Towards Conversational Visual Place Recognition},
  author={Song, Yukun and Wang, Changwei and Pei, Xingtian and Xu, Shibiao and Xu, Wenhao and Chen, Shunpeng and Zhang, Yu and Zhang, Ke and Xu, Rongtao and Feng, Xuxiang and Wang, Pengyang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={41100--41110},
  year={2026}
}

@misc{song2026dialoguevpr_arxiv,
  title={DialogueVPR: Towards Conversational Visual Place Recognition},
  author={Song, Yukun and Wang, Changwei and Pei, Xingtian and Xu, Shibiao and Xu, Wenhao and Chen, Shunpeng and Zhang, Yu and Zhang, Ke and Xu, Rongtao and Feng, Xuxiang and Wang, Pengyang},
  year={2026},
  eprint={2607.14115},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2607.14115}
}