POLIMI-ITW-S: A large-scale dataset for human activity recognition in the wild

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human activity recognition (HAR) is crucial for applications like surveillance and human-computer interaction.
• Existing HAR datasets often lack diversity and real-world conditions, limiting algorithm generalizability.
• There is a need for large-scale, in-the-wild datasets for robust HAR in public spaces.

Purpose of the Study:

• To introduce POLIMI-ITW-S, a novel, large-scale dataset for in-the-wild human activity recognition.
• To provide a comprehensive resource for developing and evaluating advanced HAR algorithms in realistic environments.
• To benchmark current state-of-the-art models on this challenging new dataset.

Main Methods:

• Collected over 46 hours of RGB video data from real shopping malls.
• Dataset comprises 22,161 video clips featuring 37 distinct activity classes.
• Annotations include person tracking bounding boxes, 2-D skeletons, and detailed activity labels.

Main Results:

• State-of-the-art HAR models achieved relatively low accuracy on the POLIMI-ITW-S dataset.
• The results highlight the challenges posed by in-the-wild activity recognition in complex environments.
• The dataset serves as a benchmark for future HAR research.

Conclusions:

• The POLIMI-ITW-S dataset presents a significant advancement for in-the-wild HAR research.
• Further research is needed to develop more robust and accurate HAR algorithms for public spaces.
• The dataset is publicly released to foster research and development in the field.