Unsupervised Action Proposals Using Support Vector Classifiers for Online Video Processing.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Action Proposals (AP) are crucial for video action understanding, enabling tasks like detection and indexing.
• Existing AP methods are supervised and require offline processing with full video access and labeled training data.
• Unsupervised and online AP methods are needed for real-world applications like robotics and video monitoring.

Purpose of the Study:

• To introduce a novel unsupervised and online approach for generating Action Proposals (AP).
• To develop a method that does not rely on labeled data or pre-trained features for learning.
• To enable efficient video processing as data arrives, suitable for real-time applications.

Main Methods:

• Utilizes a Support Vector Classifier (SVC) to identify candidate action segments from contiguous video frames.
• Employs a learning-to-rank formulation to refine and filter candidate segments based on their dynamics.
• Evaluates the approach on benchmark datasets (Thumos'14, ActivityNet) without using pre-trained features.

Main Results:

• Achieves 41% and 59% of the performance of the best supervised models on ActivityNet and Thumos'14, respectively.
• Demonstrates the efficacy of the unsupervised, online AP method on established benchmarks.
• Presents the first unsupervised approach for AP on the Thumos'14 and ActivityNet datasets.

Conclusions:

• The proposed unsupervised, online method is a viable alternative to supervised approaches for Action Proposals.
• This work opens new avenues for real-time video analysis and action understanding in resource-constrained environments.
• The approach shows promising results, validating its potential for practical applications in video monitoring and robotics.