OSFormer: One-Step Transformer for Infrared Video Small Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Small object detection in infrared videos is crucial for security and surveillance.
• Current deep learning methods struggle with temporal information and complex backgrounds.
• Existing two-step detection paradigms limit the effective use of temporal data.

Purpose of the Study:

• To introduce a novel one-step detection method for infrared video small object detection.
• To enhance the utilization of temporal information for improved accuracy.
• To develop a more efficient and effective solution compared to existing methods.

Main Methods:

• Proposed the One-Step Transformer (OSFormer) integrating a small-object-friendly transformer with a one-step detection paradigm.
• Developed the Varied-Size Patch Attention (VPA) module for adaptive feature extraction.
• Incorporated a Doppler Adaptive Filter to suppress background noise and highlight small objects.

Main Results:

• OSFormer demonstrated superior performance on the AntiUAV and InfraredUAV datasets.
• Achieved +3.1% mAP50 on AntiUAV and +4.0% mAP50-95 on InfraredUAV.
• Showcased significant reductions in parameters (-35.1%) and FLOPs (-51.0%) compared to YOLOv8-s.

Conclusions:

• OSFormer offers a highly efficient and effective solution for infrared video small object detection.
• The one-step paradigm and novel modules significantly improve detection accuracy and temporal information utilization.
• The method presents a promising advancement for surveillance and security applications.