{ "fetched_at": "2026-10-19T07:40:15Z", "source_url": "https://www.kaggle.com/competitions/ag-reid2023/data", "pages": [ { "name": "data-description", "content": "**Important:** \nBefore downloading, please read **Privacy and Usage Policy for AG-ReID dataset** at [https://www.kaggle.com/competitions/ag-reid2023/overview/privacy-policy](https://www.kaggle.com/competitions/ag-reid2023/overview/privacy-policy)\n\n# Data\n## Train\nThere will be 51,530 training images in JPG format depicting 807 identities.\n\n## Test\nThe test set comprises two subsets:\n1. `exp3_aerial_to_ground`: 808 ids | 4,348 query images | 19,259 gallery images. \n2. `exp6_ground_to_aerial`: 808 ids | 4,151 query images | 21,214 gallery images. \n\n## Annotations\nThe following table shows the annotation details for each image, including image name, person ID, camera, frame, and notes.\n|Image Name | Person ID | Camera | Frame | Note |\n| --- | --- |\n| P0001T03220A1C3F01111.jpg | P0001T03220A1 | C3 | F01111| C0: aerial, C2: wearable, C3: CCTV |\n\n## Files\n- `bounding_box_train` - contains JPG images of people for training the model.\n- `exp3_aerial_to_ground` and `exp6_ground_to_aerial` contain query and bounding_box_test for evaluating person reID model\n- `qut_attribute_v8.mat` - contains 88 binary attributes derived from 15 soft-biometric attributes for all 1615 identities. Use if required.\n- `sample_submission.csv` - contains the submission template. img_id from row 2 to row 4349 are queries of the exp3_aerial_to_ground set, and img_id from row 4350 to row 8500 are queries of the exp6_ground_to_aerial set.\n- `test.csv` - contains test template for submission.\n- `ReID2023_authorkit` - contains the template for the method description document.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "## Description\nThe AG-ReID2023 competition is an exciting opportunity for participants to showcase their expertise in person re-identification, particularly across aerial-ground environments. This novel challenge requires participants to develop effective algorithms to re-identify specific individuals across aerial and ground imagery. What sets this competition apart is the use of the new large-scale AG-ReID dataset, which comprises 100,502 frames of 1,615 identities collected using a UAV flying at multiple altitudes ranging from 15 to 45 meters, a ground-based CCTV camera, and a wearable camera on smart glasses on a university campus. When combined with other challenges in camera resolutions, occlusion, and lighting, the differences between the elevated view of aerial cameras and the horizontal view of ground cameras provide a new research area to develop practical and robust person re-identification systems.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "## How is testing done\nThe AG-ReID2023 competition evaluates submissions by computing mean Average Precision (mAP) at Rank-1. Participants' algorithms are tested on 8,499 pictures depicting 808 people who were not included in the training sample. For each picture in the test sample (query), the rest of the test dataset must be ranked, and the top result selected (the query itself should not be included in the output). For further information on submission requirements, refer to the `sample_submission.csv` file and the Dataset Description [here](https://www.kaggle.com/competitions/ag-reid2023/data). \n\nBaseline code guide can be found [here](https://github.com/agreid23/AG-ReID).\n\n## Metrics\nThis competition employs mean average precision with k=1 as the metric. For each query, the average precision is calculated on the first result, then averaged across all queries. More information on the metric can be found [here](https://towardsdatascience.com/breaking-down-mean-average-precision-map-ae462f623a52).\n\n## Submission File\nParticipants must provide a submission file containing the top-1 matching image from the `bounding_box_test` folder for each image in the test set. Each match is described using the following format:\n\n>image_id, person_id\nImageID, ImageID[13 chars]\n\n## Note\n>**IMPORTANT**: The person_id in the query set is marked and hidden from participants, which means that participants will not know the identity of the person in the query set. As a result, participants will not know whether their algorithm has correctly identified the target person at rank 1 until they submit their results and receive the evaluation feedback.\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "# Rules\n\n## General\n\n1. Your name in the leaderboard should be in the format \"Surname Name (place of study/organization).\"\n\n2. Participants must work on the problems individually.\n\n3. There is a limit of 10 attempts per day.\n\n4. Participants can only select two submissions at the end of the competition.\n\n5. Participants must exceed the medium baseline by **April 15, 2023 (9:00 AM AEST)**. If the threshold is not exceeded by the specified deadline, the participant will be excluded from the rest of the competition.\n\n6. Method description (One-page) must be submitted to [agreid2023@gmail.com](mailto:agreid2023@gmail.com) within a week after the announcement of the top 25% participants on **May 8, 2023 (9:00 AM AEST)** if you want to be included in the summary paper submitted to IJCB2023 on **May 15, 2023 (9:00 AM AEST)**.\n\n## Specific to this competition\n\n1. Only models pre-trained on ImageNet are permitted for fine-tuning.\n\n2. The usage of external data is prohibited except for the provided `qut_attribute_v8.mat` file.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Timeline", "content": "- **Feb 27, 2023 (9am AEST)** - Kaggle evaluation server opens.\n- **April 15, 2023 (9am AEST)** - Participants deadline for submissions: self-evaluation results for Part 1, software submissions for Part 2. \n- **May 8, 2023 (9am AEST)** - Announcement of the top 25% participants\n- **May 15, 2023 (9am AEST)** - Human examination and baseline results ready; paper summarising the competition submitted to IJCB. ", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Privacy Policy", "content": "Privacy and Usage Policy for AG-ReID Person ReID Dataset:\n\nBy downloading and using the AG-ReID Person ReID dataset, you agree to the following privacy and data usage policies:\n\n1. Personal Data Protection: AG-ReID dataset does not contain any personal identifying information of individuals, such as their names or any other personally identifiable information. All data in this dataset is anonymized to protect the privacy of individuals.\n2. Non-commercial Use: AG-ReID dataset is for non-commercial research and educational purposes only. Any commercial use of the dataset or any part thereof is strictly prohibited.\n3. Redistribution: Redistribution of AG-ReID dataset or any part thereof is strictly prohibited. This dataset is only available for download from [https://www.kaggle.com/competitions/ag-reid2023/data](https://www.kaggle.com/competitions/ag-reid2023/data).\n4. Attribution: Any publication or research work based on AG-ReID dataset must include proper attribution to the dataset. The recommended citation format is as follows: \"AG-ReID dataset, Kien et al., 2023\".\n5. Ethical Usage: The AG-ReID dataset should be used ethically and responsibly. Users should not use the dataset for any illegal or unethical purposes. Additionally, users should not attempt to re-identify any individuals represented in the dataset in real life. The dataset is anonymized to protect the privacy of individuals, and any attempts to re-identify individuals are strictly prohibited.\n6. Reporting: If you find any issues or problems with the AG-ReID dataset, please report them to the dataset authors or maintainers as soon as possible. \n7. No Warranty: The AG-ReID dataset is provided \"as is\" without any warranty or representation of accuracy, completeness, or reliability.\n8. Modifications: Modifications of the AG-ReID dataset are strictly prohibited without prior permission from the dataset authors or maintainers.\n9. Usage Limitation: Users should not attempt to use the AG-ReID dataset to create a competing dataset or to replicate or improve upon the dataset.\n10. Compliance: Users must comply with all applicable laws and regulations related to the use of the AG-ReID dataset, including but not limited to privacy laws and intellectual property laws.\n\nBy downloading and using the AG-ReID dataset, you agree to these terms and conditions. If you do not agree to these terms and conditions, you should not download or use the dataset.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Contact", "content": "Contact us if you have any questions at [agreid2023@gmail.com](mailto:agreid2023@gmail.com)", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }