Spatiotemporal uncertainty guided non maximum suppression for video event detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Video Event Detection (VED) is crucial for applications like autonomous driving and surveillance.
• Existing VED methods primarily focus on network architectures, neglecting uncertainty estimation.
• Uncertainty estimation is vital for reliable decision-making in VED systems.

Purpose of the Study:

• To propose an end-to-end VED neural network incorporating spatial and temporal uncertainty.
• To develop a novel Non-Maximum Suppression (NMS) method guided by estimated uncertainty, termed Spatio-Temporal Uncertainty guided NMS (STU-NMS).

Main Methods:

• Developed an end-to-end neural network for VED.
• Integrated spatial and temporal uncertainty estimation into the VED model.
• Proposed STU-NMS, which utilizes estimated uncertainty to refine detection outputs.

Main Results:

• The proposed method incorporating Spatio-Temporal Uncertainty (STU) integration outperforms existing techniques that do not model uncertainty.
• STU-NMS demonstrated improved detection performance on benchmark datasets (J-HMDB-21, UCF101-24, AVA).

Conclusions:

• Incorporating spatial and temporal uncertainty significantly enhances Video Event Detection.
• The proposed STU-NMS method offers a superior approach to VED by leveraging uncertainty for improved accuracy.