future-item/autoimagine-w2p

AutoImagine-Where2Place Dataset Card

Dataset Details

This dataset contains 100 real-world images to evaluate free space reference using spatial relations. The images are collected from various cluttered environments. Each image is labeled with a sentence describing the desired some free space and a mask of the desired region.

In particular, our dataset is an alteration of the original wentao-yuan/where2place, modified to align with the requirements of our algorithm.

Dataset Structure

• dataset folder
  • images folder
    • Contains the raw images;
  • masks folder
    • Contains the corresponding binary masks for each image;
  • point_questions.jsonl
    • Contains a list of questions asking for a single point within the desired regions. In our case, we use the autonomous imagination algorithm from our paper to find this point, instead of outputing it directly;
• prompt folder
  • Contains the prompts used for the autonomous imagination algorithm;

Resources for More Information

• Paper: https://arxiv.org/pdf/2411.18142
• Code: https://github.com/future-item/autoimagine
• Original Where2Place dataset: https://huggingface.co/datasets/wentao-yuan/where2place

Citation

If you find our work helpful, please consider citing our paper.

@article{liu2024enhancing,
  title={Enhancing visual reasoning with autonomous imagination in multimodal large language models},
  author={Liu, Jingming and Li, Yumeng and Xiao, Boyuan and Jian, Yichang and Qin, Ziang and Shao, Tianjia and Ding, Yao-Xiang and Zhou, Kun},
  journal={arXiv preprint arXiv:2411.18142},
  year={2024}
}