Investigating the Potential of Network Optimization for a Constrained Object Detection Problem.

Area of Science:

• Computer Vision
• Deep Learning
• Model Optimization

Background:

• Object detection models are typically trained on complex datasets, demanding high computational resources.
• Operational use-cases often involve constrained scenarios with fewer classes and less variance, suggesting potential for smaller models.
• Over-parameterization and transfer learning practices hinder the adoption of smaller networks in practice.

Purpose of the Study:

• To investigate the extent to which computational complexity of object detection networks can be reduced for constrained problems.
• To evaluate optimization techniques on both academic and real-world operational datasets.

Main Methods:

• Focused on the YoloV2 object detection model.
• Employed three optimization techniques: depth-wise separable convolutions, pruning, and weight quantization.
• Compared performance on a standard academic dataset (Pascal VOC) and a constrained operational dataset (LWIR person detection).

Main Results:

• The hypothesis that constrained problems allow for greater network optimization was substantiated.
• A computational complexity reduction factor of 349 was achieved on the operational dataset, compared to 9.8 on the academic dataset.
• The optimized model demonstrated a 15x speedup and a 5% increase in Average Precision (AP) on an Nvidia Jetson AGX Xavier.

Conclusions:

• Significant computational savings are achievable for object detection in constrained environments.
• The proposed optimization strategy effectively reduces model complexity without sacrificing accuracy.
• This approach enables efficient deployment of object detection in resource-limited operational settings.