- DAiSEE is a multi‑label video classification dataset.
- Size: 9,068 video snippets captured from 112 users.
- Labels: boredom, confusion, engagement, frustration.
- Created by Abhay Gupta, Arjun DCunha, KAMAL AWASTHI, Vineeth N Balasubramanian.
Human Activity Recognition (HAR) blends AI, computer vision, and data analysis to understand human motion, with kejoint annotation as the primary technique for capturing posture and activity.
- Definition: HAR merges AI, computer vision and data analysis to understand human motion.
- Method: annotate videos with kejoint labels, feed data to AI models (CNN/RNN/Transformers) for activity classification.
- Open questions: which datasets are best for training HAR models?
- Which architectures (CNNs, RNNs, Transformers) work best with kejoint data?
- How does kejoint annotation compare to skeleton or bounding‑box schemes in accuracy and labeling effort?
- What privacy and ethical concerns arise when deploying HAR in security contexts?
- Pipeline: annotation → model training → inference → insight extraction.
- Core technique: kejoint annotation captures joint positions and relationships to infer posture, activity and motion.
- Use‑cases: healthcare, sports, security – providing invaluable insights across domains.