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

<h2>Dataset Description</h2>

<p>
This dataset is developed for <strong>dog re-identification (Re-ID)</strong> across videos and images. It contains <strong>7,463 videos</strong> of <strong>1,553 distinct dog identities</strong> (averaging 4.8 videos/dog). Bounding boxes are manually annotated on the first frame of each video to support the training or fine-tuning of dog detectors.
</p>

<h3>Benchmark, Leaderboard & Baseline</h3>
<p>
To facilitate ongoing research and provide a standard starting point, we have established a public benchmark alongside this dataset. We provide a baseline model implementation, and we host an active leaderboard to track state-of-the-art methods.
</p>
<ul>
<li><strong>Leaderboard:</strong> <a href="https://project-puppies.com/leaderboard">Submit your results and view current rankings</a>.</li>
<li><strong>Baseline Implementation:</strong> Check out our <a href="https://github.com/markoMedved/DogReID-1553">GitHub repository</a> for the starter code and Re-ID pipeline.</li>
<li><strong>Paper:</strong> Detailed methodology and baseline evaluation (under review).</li>
</ul>

<h3>Data Collection</h3>
<p>
To avoid the restrictive licensing of scraped social media data, videos were crowdsourced directly from users via our <a href="https://project-puppies.com/">submission website</a> using community promotion (Reddit, Discord, Instagram) and the Prolific research platform. All submitters explicitly agreed to release their videos under the permissive <strong>Creative Commons CC0 1.0</strong> license.
</p>

<h3>Pre-processing and Anonymization</h3>
<p>
Videos were converted to <strong>MP4 (H.264)</strong>, stripped of audio, and temporally cropped so the target dog is clearly visible in the first frame.
To ensure privacy, all videos were manually reviewed. We used video editing software to permanently black out sensitive information (faces, license plates, tattoos, addresses) from the dataset.
</p>

<h3>Data Leakage Prevention</h3>
<p>
To prevent models from overfitting to environmental cues (e.g., memorizing a specific backyard instead of the dog), the gallery and query sets are <strong>scene-disjoint</strong>. Videos for each identity were manually partitioned to ensure zero background overlap between query and gallery splits.
</p>

<h3>Directory Structure</h3>

<p>Files are organized by unique UUIDs (<code>DOG_ID</code> and <code>VIDEO_ID</code>):</p>

<ul>
<li>
<strong>/Videos/</strong> — Contains the <code>.mp4</code> files.
<pre><code>/Videos/DOG_ID/DOG_ID-VIDEO_ID.mp4</code></pre>
Example:
<pre><code>/Videos/00ab4db8-3e94-44e3-a4b1-9fe316c67b60/00ab4db8-3e94-44e3-a4b1-9fe316c67b60-2cc7b112-2af8-4466-931a-c83a2466d81a.mp4</code></pre>
</li>
<li>
<strong>/Images/</strong> — Contains the <code>.jpg</code> first-frame crops.
<pre><code>/Images/DOG_ID/DOG_ID-VIDEO_ID.jpg</code></pre>
Example:
<pre><code>/Images/00ab4db8-3e94-44e3-a4b1-9fe316c67b60/00ab4db8-3e94-44e3-a4b1-9fe316c67b60-2cc7b112-2af8-4466-931a-c83a2466d81a.jpg</code></pre>
</li>
</ul>

<h3>Annotations and Metadata</h3>

<ul>
<li>
<strong>bounding_boxes.csv</strong> — Top-left coordinates and dimensions (in pixels) for the first-frame bounding boxes.
<pre><code>DOG_ID,VIDEO_ID,x_top_left,y_top_left,width,height
0062cada-a402-41bb-980e-6ae5e0672440,3f13a35c-be97-4331-b1f9-503bf4628e00,3,259,459,503</code></pre>
</li>
<li>
<strong>breeds.csv</strong> — User-submitted dog breeds.
<pre><code>DOG_ID,BREED
10541c3c-42f8-4c22-b974-5032add36da4,AlaskanMalamute</code></pre>
</li>
<li>
<strong>splits.csv</strong> — Defines the train-gallery-query splits. The <code>GROUP</code> column enforces the scene-disjoint subsets.
<pre><code>DOG_ID,VIDEO_ID,GROUP,SPLIT_CLOSED_SET,SPLIT_OPEN_SET
0062cada-a402-41bb-980e-6ae5e0672440,4f82f8da-1e3f-46f8-b48a-3a161d52cdf0,0,gallery,query</code></pre>
</li>
</ul>

<h3>Evaluation Splits Summary</h3>

<p>
The dataset is divided into two primary evaluation settings. For both, the <strong>Total number of Videos</strong> are shown first, followed by the number of unique <strong>Identities</strong> in parentheses.
</p>

<div style="margin-bottom: 20px;">
<strong>1. Closed-world Setting</strong>
<p><em>Goal: Identify a dog from a set where every query identity is guaranteed to exist in the gallery.</em></p>
<ul>
<li><strong>Train Set:</strong> 3,788 Videos (776 IDs) </li>
<li><strong>Gallery Set:</strong> 1,998 Videos (777 IDs) .</li>
<li><strong>Query Set:</strong> 1,677 Videos (680 IDs) </li>
</ul>
</div>

<div>
<strong>2. Open-world Setting</strong>
<p><em>Goal: Identify a dog while also determining if the dog even exists in the gallery at all.</em></p>
<ul>
<li><strong>Train Set:</strong> 3,708 Videos (766 IDs) </li>
<li><strong>Gallery Set:</strong> 1,296 Videos (512 IDs)</li>
<li><strong>Query Set:</strong> 2,459 Videos (777 IDs) </li>
</ul>
</div>
<h3>File Formats</h3>
<ul>
<li><strong>Videos:</strong> <code>.mp4</code></li>
<li><strong>Images:</strong> <code>.jpg</code></li>
<li><strong>Metadata, splits, and annotations:</strong> <code>.csv</code></li>
</ul>