Model compression for real-time object detection using rigorous gradation pruning

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Real-time object detection requires balancing model size and accuracy.
• Convolutional layers in deep learning models often have uneven contributions and low redundancy, complicating compression efforts.

Purpose of the Study:

• To introduce Rigorous Gradation Pruning (RGP) as an effective method for compressing object detection models.
• To maintain high detection accuracy while significantly reducing model size and increasing processing speed.

Main Methods:

• RGP utilizes a desensitized first-order Taylor approximation to evaluate filter importance for precise kernel pruning.
• The method iteratively reassesses layer significance to preserve critical layers and ensure robust detection performance.
• RGP was applied to YOLOv8 object detectors and validated on the GTSDB, Seaships, and COCO datasets.

Main Results:

• On the GTSDB dataset, RGP achieved 80% compression for YOLOv8n with a minimal 0.11% mAP0.5 drop, increasing FPS by 43.84%.
• For YOLOv8x, RGP enabled 90% compression, a 1.26% increase in mAP0.5:0.95, and a 112.66% FPS boost.
• Substantial compression rates were also observed on the Seaships and COCO datasets, confirming RGP's versatility.

Conclusions:

• Rigorous Gradation Pruning is a robust technique for compressing object detection models across various datasets.
• RGP demonstrates significant potential for developing efficient, high-speed object detection systems without compromising accuracy.