README.md

metadata

language:
  - en
license: cc-by-4.0
task_categories:
  - image-text-to-text
  - visual-question-answering
  - image-classification
pretty_name: SenseBench
tags:
  - remote-sensing
  - image-quality-assessment
  - benchmark

SenseBench

A benchmark for remote sensing low-level visual perception and description in large vision-language models.

🏠 GitHub | 📄 Paper | 🤗 Hugging Face Subset

Overview

SenseBench is the first dedicated diagnostic benchmark for remote sensing (RS) low-level visual perception and description. Driven by a physics-based hierarchical taxonomy, it features over 10K curated instances across 6 major and 22 fine-grained RS degradation categories. It is designed to evaluate whether Vision-Language Models (VLMs) can overcome the domain gap to perceive and articulate RS-specific artifacts.

The benchmark evaluation consists of two complementary protocols:

1. Objective low-level visual perception: Evaluating the model's ability to identify the presence and type of distortions.
2. Subjective diagnostic description: Evaluating the model's ability to articulate RS artifacts in natural language based on completeness, correctness, and faithfulness.

Supported Tasks

• Visual Question Answering: Multiple-choice questions assessing degradation type and severity.
• Image-to-Text / Diagnostic Description: Natural language generation describing visual artifacts.

Language

• English

Data format

Each example contains image paths, a question, an answer, and metadata describing the distortion type.

{
  "id": "4fda312e-70d2-4df7-b1f7-2f06955bf338",
  "images": [
    "images/4fda312e-70d2-4df7-b1f7-2f06955bf338_0.png",
    "images/4fda312e-70d2-4df7-b1f7-2f06955bf338_1.png"
  ],
  "question": "Using the options provided, rate the overall quality of Image 2 compared to Image 1.
A.No/Slight distortion
B.Moderate distortion
C.Severe distortion",
  "answer": "A",
  "meta": {
    "image_count": "multi",
    "modality": "RGB",
    "task": "how",
    "domain": "general",
    "distortion_family": "blur",
    "distortion_type": "blur_gaussian",
    "distortion_complexity": "single",
    "comparison": "intra-image"
  }
}