WILDTRACK pedestrian detection dataset

WILDTRACK is a dataset of video recordings of over one thousand students, faculty and others to-ing and fro-ing 'in the wild' outside the main building at ETH University, Zurich. 

Developed by École Polytechnique Fédérale de Lausanne and ETH University researchers in Switzerland, the dataset comprises seven 35-minute videos captured by seven high-definition GoPro cameras during good weather conditions.

The videos were subsequently annotated using Mechanical Turk to mark the positions of people in each frame.

Risks and harms 🛑

The WILDTRACK pedestrian detection dataset has been accused of violating privacy and enabling potentially unethical research and applications in China and elsewhere.

Transparency and accountability 🙈

The VGG Face dataset is seen to suffer from several transparency limitations.

• Data collection methodology. There's limited information publicly available about how the images were collected and curated. This lack of transparency makes it difficult to assess potential biases in the data gathering process.

• Demographic representation. The dataset may not have a balanced representation across different demographics such as age, gender, ethnicity, and geographical regions. Without clear documentation of the demographic breakdown, it's challenging to evaluate the dataset's diversity and potential biases.

• Consent and privacy. It is unclear how consent was obtained for the individuals whose images are included in the dataset, raising ethical questions about privacy and data usage rights.

• Annotation process. Details about the annotation process, including how identities were verified and labeled, are not fully disclosed, affecting the reliability of the ground truth labels.

Research, advocacy 🧮

• Harvey, A., LaPlace, J. (2019). Exposing.ai

• Zhu H., Qi Y., Shi H., Li N., Zhou H. (2018). Human Detection Under UAV: an Improved Faster R-CNN Approach

• Chakraborty I., Hua G. (2019). Priming Deep Pedestrian Detection with Geometric Context