FBD-SV-2024: Flying Bird Object Detection Dataset in Surveillance Video.

Area of Science:

• Computer Vision
• Machine Learning
• Wildlife Monitoring

Background:

• Accurate detection of flying birds in surveillance footage is crucial for various applications, including wildlife management and public safety.
• Existing datasets may not adequately represent the complexities of real-world surveillance scenarios, such as small object sizes and varied appearances.

Purpose of the Study:

• To introduce the Flying Bird Dataset for Surveillance Videos (FBD-SV-2024), a novel resource for training and benchmarking flying bird detection algorithms.
• To provide a challenging benchmark that reflects realistic surveillance conditions for evaluating algorithm performance.

Main Methods:

• The FBD-SV-2024 dataset was curated, comprising 483 video clips with 28,694 frames, featuring 28,366 annotated flying bird instances.
• The dataset captures birds in realistic surveillance settings, highlighting challenges like inconspicuous features, small sizes, and shape variability.

Main Results:

• Experiments were conducted using state-of-the-art video object detection algorithms on the FBD-SV-2024 dataset.
• The results indicate that even advanced algorithms struggle to achieve optimal performance on this challenging dataset, underscoring its difficulty.

Conclusions:

• The FBD-SV-2024 dataset presents a significant challenge for current flying bird detection technologies.
• This dataset will drive the development of more robust and accurate algorithms for bird detection in surveillance applications.