MOInst Dataset (Multi-Object Instruction)

Overview

MOInst (Multi-Object Instruction) is a specialized dataset created for training and evaluating computer vision models on their ability to perceive overlooked information in images. The dataset is built by carefully selecting and annotating images from the Visual Genome dataset, focusing on object instances that are typically missed by standard vision models.

Relation to GiVE

This dataset was developed as part of the research presented in "GiVE: Guiding Visual Encoder to Perceive Overlooked Information" (ICME 2025). GiVE is a framework designed to enhance vision models by guiding them to perceive information that might be overlooked. The MOInst dataset serves as both a training resource and evaluation benchmark for this purpose.

Dataset Structure

The dataset is organized into:

• Training set: MOInst_train.json - Contains annotated images for model training
• Test set: MOInst_test.json - Used for evaluation

Data Source

All images and captions in the MOInst dataset are sourced from the Visual Genome dataset. We further annotated the focused objects.

Usage

The MOInst dataset is designed for:

1. Training vision models to better perceive overlooked information in images
2. Evaluating model performance on identifying missed object instances
3. Benchmarking vision-language models on fine-grained visual understanding tasks

Citation

If you use this dataset in your research, please cite the GiVE paper:

@article{DBLP:journals/corr/abs-2410-20109,
  author       = {Junjie Li and
                  Jianghong Ma and
                  Xiaofeng Zhang and
                  Yuhang Li and
                  Jianyang Shi},
  title        = {GiVE: Guiding Visual Encoder to Perceive Overlooked Information},
  journal      = {CoRR},
  volume       = {abs/2410.20109},
  year         = {2024},
  url          = {https://doi.org/10.48550/arXiv.2410.20109},
  doi          = {10.48550/ARXIV.2410.20109},
  eprinttype    = {arXiv},
  eprint       = {2410.20109},
  timestamp    = {Thu, 28 Nov 2024 21:32:45 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2410-20109.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}