RSTPReid (Real Scenario Text-based Person Re-identification)

为了正确处理真实场景，我们基于 MSMT17 [2] 构建了一个名为 Real Scenario Text-based Person Re-identification (RSTPReid) 的新数据集。 RSTPReid 包含来自 15 个摄像头的 4,101 人的 20505 张图像。每个人有 5 张对应的图像，由不同的相机在不同时间段内拍摄，具有复杂的室内外场景变换和背景，这使得 RSTPReid 更具挑战性，更能适应真实场景。每张图片都带有 2 个文字描述。对于数据划分，分别使用 3701（索引 < 18505）、200（18505 <= 索引 < 19505）和 200（索引 >= 19505）身份进行训练、验证和测试（由 json 文件中的项目“拆分”标记）。每个句子不少于 23 个单词。
考虑到 RSTPReid 数据集是新建的，与 CUHK-PEDES 相比数据量较小，我们将每个模型分别训练了 10 次，并在表 1 中报告了验证集上的平均结果，以避免偶尔出现不稳定的性能。此外，在本文中，我们仅使用 RSTPReid 对提出的五种对齐范式进行了初步分析，并将结果作为每个组件有效性的进一步证明。本文中的大部分实验仍然使用 CHUK-PEDES 进行，其中报告了测试集的最终结果。
在 RSTPReid 的测试集上，DSSL 分别达到了 top-1、top-5 和 top-10 的 39.05%、62.60% 和 73.95% 的准确率。将与 CUHK-PEDES 一起对 RETPReid 进行更详细的实验分析，这将在本文的扩展部分中报告。

To properly handle real-world scenarios, we construct a new dataset named Real Scenario Text-based Person Re-identification (RSTPReid) based on MSMT17 [2]. RSTPReid contains 20,505 images of 4,101 individuals from 15 cameras. Each individual has 5 corresponding images captured by different cameras across different time periods, featuring complex indoor-outdoor scene transitions and backgrounds, which makes RSTPReid more challenging and better adapted to real-world scenarios. Each image is paired with 2 textual descriptions. For data splitting, 3,701 identities (index < 18505), 200 identities (18505 <= index < 19505), and 200 identities (index >= 19505) are used for training, validation, and testing respectively, as marked by the "split" item in the json file. Each sentence contains no fewer than 23 words.

Considering that RSTPReid is a newly built dataset with a smaller scale compared to CUHK-PEDES, we train each model 10 times separately and report the average results on the validation set in Table 1 to avoid occasionally unstable performance. Additionally, in this paper, we only use RSTPReid to conduct a preliminary analysis of the five proposed alignment paradigms, and take the results as further evidence of the effectiveness of each component. Most experiments in this paper are still conducted using CHUK-PEDES, where the final results on the test set are reported.

On the test set of RSTPReid, DSSL achieves accuracies of 39.05%, 62.60%, and 73.95% for top-1, top-5, and top-10 respectively. More detailed experimental analyses will be conducted on RETPReid together with CUHK-PEDES, which will be reported in the extended version of this paper.

搜集汇总

背景与挑战

背景概述

RSTPReid是一个基于文本的人员重识别数据集，专注于真实场景应用。它包含4,101人的20,505张图像，来自15个摄像头，每人有5张图像，具有复杂的室内外场景变换和背景，每张图像配有2个文字描述，用于训练和评估文本检索模型。数据集由北京航空航天大学等机构于2021年发布，旨在提升模型在真实环境中的适应性和挑战性。