FSC147

该数据集名为FSC147，是一个包含147个对象类别的3次射击计数数据集，适用于进行类别无关的计数基准测试。该数据集包含6135张图像，其中3659张用于训练。每张图像都标注了3个示例边界框和一个对象密度图。此外，训练集、验证集和测试集的类别是互斥的，总共包含89个训练类别、29个验证类别和29个测试类别。该数据集的规模为3次射击计数，任务类型为少样本类别无关计数。

The dataset named FSC147 is a 3-shot counting dataset encompassing 147 object categories, designed for category-agnostic counting benchmarking. It contains a total of 6,135 images, among which 3,659 are allocated for training. Each image is annotated with three exemplary bounding boxes and an object density map. Furthermore, the categories in the training, validation and test sets are mutually exclusive, with 89 training categories, 29 validation categories and 29 test categories respectively. This is a 3-shot counting dataset with its task type being few-shot category-agnostic counting.

搜集汇总

背景与挑战

背景概述

FSC147是一个用于物体计数任务的数据集，包含图像和预计算的密度图，支持训练和评估计数模型。该数据集是CVPR 2021论文《Learning To Count Everything》的实现基础，提供了快速演示、评估和训练代码。