Foreground separation knowledge distillation for object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep learning models are essential for computer vision but resource-intensive, hindering deployment on edge devices.
• Knowledge distillation (KD) is a key technique for model compression, but existing methods struggle with noise in object detection.
• Current KD approaches for object detection often overlook noise or inadequately separate foreground/background, impacting student model accuracy.

Purpose of the Study:

• To introduce a novel foreground separation distillation (FSD) method for efficient object detection model compression.
• To enhance student model accuracy by minimizing irrelevant information and maximizing knowledge transfer from teacher models.
• To improve the deployability of deep learning models on resource-constrained devices.

Main Methods:

• Proposed Foreground Separation Distillation (FSD) utilizing Gaussian heatmaps for foreground-background distinction.
• Implemented a method to extract channel features by converting spatial feature maps into probabilistic forms.
• Applied FSD to the YOLOX object detection framework.

Main Results:

• The YOLOX detector enhanced with FSD demonstrated superior performance on fall detection and VOC2007 datasets.
• Achieved 73.1% mean average precision (mAP) on the Fall Detection dataset, a 1.6% improvement over the baseline.
• FSD effectively reduced irrelevant information and improved knowledge utilization from the teacher model.

Conclusions:

• Foreground Separation Distillation (FSD) offers a significant advancement in knowledge distillation for object detection.
• FSD enhances model accuracy and efficiency, making deep learning models more suitable for resource-limited environments.
• The proposed method provides a robust solution for noise reduction and effective knowledge transfer in model compression.