Vision-Judge

Overview

Vision-Judge is a benchmark dataset for evaluating the judging ability of Multimodal Large Language Models (MLLMs) on diverse vision tasks. Given a visual input and multiple model outputs, an MLLM acting as a judge is asked to rank the outputs by quality. Vision-Judge covers 10 vision task categories, enabling systematic evaluation of whether MLLMs can accurately assess and compare the outputs of specialized vision models.

Dataset Structure

This repository contains two components:

1. Preview split (train): A curated sample of ~235 examples (20 per task, with all generation subtask examples included) for quick exploration via the Dataset Viewer.
2. Full annotations (annotations/): Complete annotation JSON files for all tasks.
3. Full images (images/): Complete image JSON files for all tasks.

Preview Split Columns

Column	Type	Description
task	string	Vision task category (e.g., depth, keypoint, generation_t2i)
image_id	string	Unique identifier for the source image
error_type	string	Type of error present in model outputs (null for model-based tasks)
instance_type	string	Instance category (null for non-instance tasks)
coi	list	Category of interest for the evaluation
prompt	string	The evaluation prompt shown to the judge model
media	images list	Input image(s) for the task
annotation_ids	list	IDs of the model output annotations being ranked

Task Categories

Task	Description
depth	Monocular depth estimation
generation	Image generation (text-to-image, inpainting, controllable generation)
instance_segmentation_model	Instance segmentation with real model outputs
instance_segmentation_synthetic	Instance segmentation with synthetic error injection
keypoint	Human keypoint detection
lowlevel	Low-level vision (super-resolution, deblurring, desnowing, deraining)
object_detection	Object detection and bounding box prediction
referring_segmentation_model	Referring expression segmentation with real model outputs
referring_segmentation_synthetic	Referring expression segmentation with synthetic errors
semantic_segmentation	Semantic scene segmentation

Ranking and Scoring Notes

Rankings in Vision-Judge are based on task-specific quality metrics (e.g., IoU for segmentation tasks). Tied scores are not included as comparative pairs in evaluation prompts. For example, in instance segmentation, outputs where all models produce IoU = 0 (wrong instance predictions) are not directly compared against each other — only comparisons involving at least one output with IoU > 0 are included. This ensures that ranking judgments are meaningful and grounded in measurable quality differences.

Usage

from datasets import load_dataset

# Load the preview split
ds = load_dataset("Anonymous-44kk/Vision-Judge", split="train")

# Filter by task
depth_examples = ds.filter(lambda x: x["task"] == "depth")

License

This dataset is released under CC BY 4.0.