__ The dataset is provided for research purposes only. To obtain access to the data, a [license agreement](https://tuni-my.sharepoint.com/:b:/g/personal/jenni_raitoharju_tuni_fi/EWtLHFXv_29JhPQJmG3VGbcBvz348ulYXw95eaIe61EU2A?e=ZOg8ac) must be signed and sent to jenni.raitoharju at tuni.fi__ .
The dataset contains 96 high resolution RGB images of multiple people with all of their locations annotated for each image. The dataset is meant to be used for evaluating the performance of methods that can automatically estimate social distances between people from RGB images.
All of the images are in JPG format. 81 of the images have the resolution 4180x2768 and 15 of the images have the resolution 4080x2720. Two different camera models were used and the sensor size for both of these cameras is 36 mm in width and 24 mm in height. The camera models are Canon EOS 5D Mark II and Canon EOS 6D Mark II. Different focal length lenses were used for capturing the images. These focal lengths are: 16, 24, 35, 50, 105, 200, and 300 mm. Focal length information is present in the metadata of each image. The cameras were stabilized on a tripod and the tripod’s height was 135 cm for all images.
The dataset contains images from 2 different photoshoots/locations that were taken in an outdoor and an indoor setting. During each of the photo shoots, test subjects stayed on the same known positions, while the photographer changed his position and used multiple cameras and lenses at each spot.
The dataset contains annotations for the relative ground truth 3D locations of each person and the cameras and also contains pixel locations for 4 body parts of each person (center of eyes, torso, shoulders and head). Along with the dataset, we provide codes to evaluate our own social distance estimation method and we also provide the necessary API for other methods to evaluate their performances on this dataset. Furthermore, additional images to the dataset can be added by following the annotation format. Detailed explanations are found in the attached README.txt file.
All of the evaluation code and the tutorial may also be found in: [https://github.com/mertseker-dev/social-distance-estimation](https://github.com/mertseker-dev/social-distance-estimation).
M. Seker, A. Mannisto, A. Iosifidis, J. Raitoharju, Mert Seker, Anssi Männistö, Alexandros Iosifidis, Jenni Raitoharju, __ Automatic Social Distance Estimation From Images: Performance Evaluation, Test Benchmark, and Algorithm__ , arXiv:2103.06759, 2021. Show more
COVID-19, human pose estimation, object detection, person detection, social distance estimation, social distancing