Scalable Semi-Automatic Annotation for Multi-Camera Person Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Video Analysis

Background:

• Evaluating multi-camera people-tracking systems requires extensive and reliable position annotations.
• Manual annotation is time-consuming and labor-intensive, hindering scalability for large video datasets.

Purpose of the Study:

• To propose a generic, semi-automatic methodology for generating reliable position annotations for multi-camera people-tracking evaluation.
• To reduce the manual effort required for annotating large-scale video data.

Main Methods:

• Automated generation of annotations by estimating consensus tracking results from multiple trackers and detectors.
• Classification of automated annotations as reliable or unreliable.
• Human verification of a small subset of unreliable tracks using a binary decision task.

Main Results:

• 80% of video frames were automatically annotated with a tracking accuracy of 60 cm.
• An additional 20% of frames were annotated with approximately 2.4 hours of manual labor.
• 99% accuracy was achieved for automatically annotated frames after visual inspection.

Conclusions:

• The proposed semi-automatic methodology significantly enhances the efficiency of generating reliable annotations for people-tracking evaluation.
• The framework is generic, adaptable to new datasets, and supports additional trackers.
• Exploratory study on the multi-target case demonstrates broader applicability.