README.md

metadata

task_categories:
  - image-segmentation
tags:
  - instance-segmentation
  - instruction-following

Inst2Seg

Inst2Seg is a high-quality and large-scale instruction-based instance segmentation dataset and benchmark introduced in the paper InstructSAM: Segment Any Instance with Any Instructions.

The dataset couples free-form instructions with instance-level masks to support training and evaluation for unified instruction-driven segmentation. It covers several types of instructions:

• Category prompts
• Referring expressions
• Reasoning-style instructions

Project Resources

• Paper: InstructSAM: Segment Any Instance with Any Instructions
• Code: GitHub - InstructSAM

Dataset Description

Inst2Seg is designed to address the challenges of multi-instance segmentation under arbitrary natural-language instructions. It enables models to perform high-level instruction understanding and instance-level set prediction. Unlike traditional semantic segmentation, which focuses on regions, Inst2Seg requires models to resolve compositional reasoning and enumerate specific instances based on the provided instructions.

Data Structure and Usage

According to the official repository, the dataset typically consists of annotation files in JSON format. To use the dataset:

1. Download the training and evaluation annotation JSON files from the project's data repository.
2. Download the raw images from the official sources of the original datasets (e.g., COCO, RefCOCO, etc.).
3. Place the JSON files under data/training and data/eval as specified in the setup instructions.

Evaluation can be performed using the provided scripts:

bash evaluation/scripts/eval_inst2seg.sh
bash evaluation/scripts/eval_reasonseg.sh
bash evaluation/scripts/eval_grefcoco_ap.sh
bash evaluation/scripts/eval_roborefit.sh

Citation

If you find this dataset or the associated framework useful, please cite:

@article{yuan2026instructsam,
  title     = {InstructSAM: Segment Any Instance with Any Instructions},
  author    = {Yuqian Yuan, Wentong Li, Zhaocheng Li, Yutong Lin, Juncheng Li, Siliang Tang, Jun Xiao, Yueting Zhuang, Wenqiao Zhang},
  year      = {2026},
  journal   = {arXiv},
}