Awesome

DDFAV: Remote Sensing Large Vision Language Models Dataset and Evaluation Benchmark

This repository contains the instruction set, evaluation method files and corresponding pictures I made.

DDFAV Instruction Set

This is an example of an instruction set I made. The rest of the instruction sets and corresponding pictures are in the DDFAV_Instruction Set folder.

Each remote sensing image instruction set generates 8 questions, including a detailed description of the image, a complex question reasoning, two visual question answering tasks for color, two visual question answering tasks for counting, and two visual question answering tasks for object location.

RSPOPE Hallucination Assessment Method

This is an example of the RSPOPE hallucination evaluation method I made. The rest of the evaluation files and corresponding pictures are in the rspope_evaluation folder.

There are 9 settings (easy, meidum, hard) and (random, popular, adversarial) in total. Based on the original POPE random, popular, adversarial settings, the easy setting requires at least 2 different types of objects, the number of objects is within 5, and each image contains 6 binary classification problems; the medium setting requires at least 3 different types of objects, the number of objects is between 6 and 10, and each image contains 8 binary classification problems; the hard setting requires at least 4 different types of objects, the number of objects is more than 11, and each image contains 10 binary classification problems.

DDFAV Dataset

The DDFAV dataset uses 5 remote sensing datasets including (DIOR, DOTA, FAIR1M, VisDrone-2019, AI-TOD).