DogReID-1553: A Large-Scale Dog Re-Identification Video Dataset and Benchmark

Dataset Description This dataset is developed for research in dog re-identification (Re-ID), where the goal is to recognize individual dogs across Videos and Images. It includes multiple videos per dog identity, corresponding images extracted from the first frame of each video, and text descriptions generated from those images in JSON format using the GPT-5 model. The dataset contains 1,553 distinct dog identities and 7,463 videos, averaging 4.8 videos per identity. Bounding boxes were manually annotated around target dogs in the first-frame images to support localization-based training and evaluation. In addition to the dataset we also added a leaderboard on our website. You can submit your results there. For more information you can also read our paper (TODO). Directory Structure and Identifier Format The dataset is organized into two main folders — Videos and Images— each following an identical structure based on unique dog identities. Each dog is represented by a unique identifier (DOG_ID) in UUID format (e.g., 00ab4db8-3e94-44e3-a4b1-9fe316c67b60), and all files related to that identity are grouped within a subfolder named after its DOG_ID. /Videos/ — Each DOG_ID subfolder contains one or more video files of the same dog. Each video has its own unique VIDEO_ID and follows the naming pattern: /Videos/DOG_ID/DOG_ID-VIDEO_ID.mp4 Example: /Videos/00ab4db8-3e94-44e3-a4b1-9fe316c67b60/00ab4db8-3e94-44e3-a4b1-9fe316c67b60-2cc7b112-2af8-4466-931a-c83a2466d81a.mp4 /Images/ — Contains the first-frame image extracted from each video. Images follow the same hierarchical structure and naming pattern as the videos: /Images/DOG_ID/DOG_ID-VIDEO_ID.jpg bounding_boxes.csv — Contains coordinates for manually annotated bounding boxes: DOG_ID,VIDEO_ID,x_top_left,y_top_left,width,height x_top_left, y_top_left — pixel coordinates of the bounding box’s top-left corner. width, height — bounding box dimensions in pixels. splits.csv — Define dataset partitions for closed-world and open-world evaluation. In addition to that we addded the coloumn GROUP which determines the space that the video was taken. For more details you can read our paper at TODO. Each row specifies whether a video belongs to thetrain, query or gallery set for open as well as closed set: DOG_ID,VIDEO_ID,GROUP,SPLIT_CLOSED_SET,SPLIT_OPEN_SET 0062cada-a402-41bb-980e-6ae5e0672440,4f82f8da-1e3f-46f8-b48a-3a161d52cdf0,0,gallery,query breeds.csv — Contains metadata describing the breed of each dog: DOG_ID,VIDEO_ID,BREED 44833f04-7e94-4e7b-9549-4a0b19cd4f5f,e16c3a9f-df3b-46da-b211-e04515e0df44,Pitlabmix File Formats Videos: .mp4 Images: .jpg Metadata, splits, and annotations: .csv数据集说明 本数据集专为犬只重识别（Re-identification, Re-ID）研究开发，核心目标为实现跨视频与图像场景的个体犬只精准识别。数据集包含每个犬只身份对应的多段视频、从每段视频首帧提取的对应图像，以及依托这些图像由GPT-5模型生成的JSON格式文本描述。本数据集涵盖1553个独立犬只身份与7463段视频，平均每个身份对应4.8段视频。研究人员已针对首帧图像中的目标犬只完成人工边界框标注，以支持基于定位的模型训练与评估。此外，我们在官方网站增设了排行榜供研究者提交实验结果，更多细节可参阅我们的论文（待补充）。 目录结构与标识符格式 本数据集分为两大核心文件夹——Videos（视频文件夹）与Images（图像文件夹），二者均基于唯一犬只身份采用统一的层级组织结构。每只犬只对应一个UUID格式的唯一标识符（DOG_ID，例如00ab4db8-3e94-44e3-a4b1-9fe316c67b60），该犬只相关的所有文件均存储在以其DOG_ID命名的子文件夹中。 /Videos/ — 每个DOG_ID子文件夹包含对应犬只的一段或多段视频。每段视频拥有独立的VIDEO_ID，文件命名遵循如下格式：/Videos/DOG_ID/DOG_ID-VIDEO_ID.mp4 示例：/Videos/00ab4db8-3e94-44e3-a4b1-9fe316c67b60/00ab4db8-3e94-44e3-a4b1-9fe316c67b60-2cc7b112-2af8-4466-931a-c83a2466d81a.mp4 /Images/ — 存储从每段视频首帧提取的图像文件，其层级结构与命名规则与视频文件夹完全一致：/Images/DOG_ID/DOG_ID-VIDEO_ID.jpg bounding_boxes.csv — 包含人工标注的边界框坐标，字段依次为：DOG_ID,VIDEO_ID,x_top_left,y_top_left,width,height 其中x_top_left与y_top_left为边界框左上角的像素坐标，width与height为边界框的像素尺寸。 splits.csv — 定义用于闭世界与开世界评估的数据集划分规则。本文件新增GROUP字段，用于标识视频的采集场景。更多细节可参阅我们的论文（待补充）。文件每行指定一段视频所属的分组，格式为：DOG_ID,VIDEO_ID,GROUP,SPLIT_CLOSED_SET,SPLIT_OPEN_SET 示例行：0062cada-a402-41bb-980e-6ae5e0672440,4f82f8da-1e3f-46f8-b48a-3a161d52cdf0,0,gallery,query breeds.csv — 包含各犬只的品种元数据，字段依次为：DOG_ID,VIDEO_ID,BREED 示例行：44833f04-7e94-4e7b-9549-4a0b19cd4f5f,e16c3a9f-df3b-46da-b211-e04515e0df44,Pitlabmix 文件格式说明 视频文件格式：.mp4 图像文件格式：.jpg 元数据、数据集划分与标注文件格式：.csv