Welcome to DataCV2024!
This is the development kit repository for the 2nd CVPR DataCV Challenge. Here you can find details on how to download datasets, run baseline models, and evaluate the performance of your model. Evaluation can be performed remotely on the CodeLab evaluation server. Please also refer to the main website for competition details, rules, and dates.
Have fun!
This year’s challenge focuses on Training set search for object detection. We consider a scenario where we have access to the target domain, but cannot afford on-the-fly training data annotation, and instead would like to construct an alternative training set from a large-scale data pool such that a competitive model can be obtained. Specifically, we focus on vehicle detection. We newly introduce Regin100 as our target and reuse the existing 7 well-known datasets as a source pool.
The competition will take place from Feb -- Apr 2024, and the top-performing teams will be invited to present their results at the workshop at CVPR 2024 in Jun, Seattle.
Please see the code from the top teams in the 2024 DataCV Challenge (not in any particular order)
Source pool: [GoogleDrive] and [BaiduYun] (pwd: jp2x)
Region100 training and testA: [GoogleDrive] and [BaiduYun] (pwd: 4ls8)
TestA annotation [GoogleDrive] and [BaiduYun] (pwd: krd6)
Region100 testB [GoogleDrive] and [BaiduYun] (pwd: n5jv)
The target dataset: Region100 benchmark consists of footage captured by 100 static cameras from various regions in the real world. For videos from each different region, the first 70% is used for model training, while the remaining 30% is designated for validation and testing.
The target domain consists of real-world images from 100 regions in the world. In total, there are 21871 images. We have provided region camera index information for both source and target training sets.
- For the training set, it contains 15,368 images.
- For the test A set, it contains 2,134 images.
- For the test B set, it contains 4,368 images. The test B set will be reserved for the final ranking.
The source pool comprises datasets from seven existing sources: ADE, BDD, Cityscapes, COCO, VOC, Detrac, and KITTI. We have standardized the labeling format to match that of COCO, and filtered out labels that are not related to vehicles. In total, the collection contains 176,491 images.
The challenge dataset split is organized as follows:
├── Region100/
(Source pool from existing benchmarks)
│ ├── source_pool/ /* source training images
│ ├── voc_train/
│ ├── VOC2007/
| ├── VOC2012
│ ├── kitti_train/
│ ├── 000000.png
| ├── 000001.png
| ...
│ ├── coco_train/
│ ├── 000000000009.jpg
| ├── 000000000025.jpg
| ...
│ ├── cityscapes_train/
│ ├── aachen_000000_000019_leftImg8bit.jpg
| ├── aachen_000001_000019_leftImg8bit.jpg
| ...
│ ├── bdd_train/
│ ├── 0a0a0b1a-7c39d841.jpg
| ├── 0a0b16e2-93f8c456.jpg
| ...
│ ├── ade_train/
│ ├── ADE_train_00000001.jpg
| ├── ADE_train_00000002.jpg
| ...
| ├── detrac_train/
│ ├── voc_annotation.json
│ ├── kitti_annotation.json
│ ├── coco_annotation.json
│ ├── cityscapes_annotation.json
│ ├── bdd_annotation.json
│ ├── ade_annotation.json
(Target region100 dataset collected from real world)
│ ├── train/ /* training set
│ ├── 001/
| ├── 001_000001.jpg
| ├── 001_000002.jpg
| ...
│ ├── 002/
| ├── 002_000001.jpg
| ├── 002_000002.jpg
| ...
│ ├── testA/ /* test A set
│ ├── 001_000001.jpg
│ ├── 001_000002.jpg
│ ├── 001_000003.jpg
│ ├── 001_000004.jpg
| ...
│ ├── testB/ /* test B set
│ ├── 001_000001.jpg
│ ├── 001_000002.jpg
│ ├── 001_000003.jpg
│ ├── 001_000004.jpg
| ...
The aim of this challenge is to find no more than 8000 images from the source pool to train a competent model for our region100 target. A baseline algorithm for training set search has been provided.
The first three digits correspond to the region code, and the following six digits correspond to the image number within the region. For example, 001_000001.jpg is the first image captured from the first region camera. "001" is the region code, and "000001" is the image counting number.
By downloading these datasets you agree to the following terms:
- The data is to be utilized solely for research and educational purposes that do not seek profit.
- The distribution of images obtained from the data is prohibited.
- The providers of the data disclaim all guarantees, including but not limited to those concerning the violation of rights or the data's suitability for a specific use.
- Users assume all responsibility for their application of the data.
To manage the Python code, it is recommended to install Anaconda.
For creating an environment,
conda create -n tss python=3.7 -y
conda activate tssBesides, you will need to install pytorch 1.12, please modify cuda version according to the version installed in the system.
pip install torch==1.12.0+cu116 torchvision==0.13.0+cu116 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu116The installation of mmcv library.
pip install -U openmim
mim install mmcv-full==1.7.0
pip install yapf==0.40.1The installation of mmdetction.
cd mmdetection/
pip install -v -e . Additionally
pip install scikit-learn
pip install scipy==1.2.1We implement the SnP framework proposed by Yao etal. We have adapted the SnP framework to the detection task, the this repo will create a searched training set with 8,000 images.
Unlike traditional challenges where the training set is fixed, allowing participants to optimize their model or algorithm, in our 'training set search' challenge, we keep the model or algorithm fixed and allow participants to optimize the training set. We fix our task model as RetinaNet, which is available on this link.
If you find any bugs please open an issue.
@inproceedings{yao2023large,
title={Large-scale Training Data Search for Object Re-identification},
author={Yao, Yue and Lei, Huan and Gedeon, Tom and Zheng, Liang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={15568--15578},
year={2023}
}