MIT place2 dataset 场景图像数据集

Places2 是一个场景图像数据集，包含 1千万张 图片，400多个不同类型的场景环境，可用于以场景和环境为应用内容的视觉认知任务。 该目标是识别照片中描绘的场景类别。此任务的数据来自Places数据集，其中包含属于400多个唯一场景类别的 1000多万个图像。具体而言，数据将被分为8百万个用于训练的图像，用于验证的36K图像和用于来自365个场景类别的328K用于测试的图像。请注意，每个类别的训练图像分布不均匀，范围从4,000到40,000，模仿场景出现的更自然的频率。 对于每个图像，算法将按置信度的降序生成最多5个场景类别的列表。标签的质量将根据最符合图像地面实况标签的标签进行评估。这个想法是允许算法识别图像中的多个场景类别，因为许多环境具有多个标签（例如，酒吧也可以是餐馆），并且人们经常用不同的词语描述一个地方（例如森林小径，森林，树木）。

Places2 is a scene image dataset consisting of over 10 million images across more than 400 distinct scene categories, designed for visual cognition tasks centered on scenes and environments. The core objective of this dataset is to identify the scene category depicted in a photograph. The data for this task is derived from the Places dataset, which encompasses over 10 million images belonging to over 400 unique scene categories. Specifically, the dataset is partitioned into 8 million training images, 36,000 validation images, and 328,000 test images spanning 365 scene categories. It should be noted that the distribution of training images per category is uneven, ranging from 4,000 to 40,000, which mimics the realistic natural frequency of scene occurrences in real life. For each input image, algorithms are required to generate a list of up to 5 scene categories sorted in descending order of confidence. The quality of the predicted labels will be evaluated based on the label that best matches the ground-truth label of the image. This design rationale is to allow algorithms to recognize multiple scene categories within a single image, as many environments have multiple applicable labels (e.g., a bar can also be classified as a restaurant), and people often use diverse terms to describe the same location (e.g., forest trail, forest, trees).

搜集汇总

背景与挑战

背景概述

MIT place2 dataset是一个大规模场景图像数据集，包含1千万张图片，覆盖400多个不同类型的场景环境，适用于视觉认知任务。数据集分为训练、验证和测试集，来自365个场景类别，可用于场景识别和分类等应用。