CountBench

该数据集是一个新的图像文本计数基准，旨在评估模型对物体计数理解的能力。它包含了540张精心挑选且类别平衡的图片。该数据集能够对计数准确性进行基准测试，未经训练的情况下，以CLIP B/32模型为基础的性能准确度为27.5%。规模上，数据集包含了540张图片，任务是对图像中的物体进行计数。

This dataset is a novel image-text counting benchmark designed to evaluate models' ability to understand object counting. It contains 540 carefully selected and class-balanced images. This benchmark enables the evaluation of counting accuracy, with the CLIP B/32 model achieving an accuracy of 27.5% without prior training. Regarding its scale, the dataset includes 540 images, and the task is to count the objects in the images.

搜集汇总

背景与挑战

背景概述

CountBench是一个用于训练和验证CountCLIP模型的数据集，包含约2000张计数图像和13000张非计数图像，旨在提升视觉语言模型在零样本计数任务中的准确性。