What it is
Benchmark for instruction-conditioned image segmentation across explicit vs implicit wording, target count, and four visual domains. 4,800 image-text-mask samples organized as 1,200 human-reviewed paired units across 4 domains, crossing explicit/implicit instructions with single/multiple targets.
Key results
- 4,800 instruction-level samples
- 1,200 paired evaluation units across 4 domains
- Image + English instruction + COCO RLE mask; Apache-2.0
Useful for evaluating segmentation models on explicit vs implicit referring instructions across diverse image domains.