Instance mask alignment for object detection knowledge distillation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Knowledge distillation is a key technique for improving object detection models.
• A significant performance gap often exists between teacher and student models, especially with heterogeneous detector types.

Purpose of the Study:

• To propose a novel knowledge distillation framework, Instance Mask Alignment (IMA), to bridge the teacher-student performance gap in object detection.
• To enhance student model performance by leveraging instance mask information and cascade alignment modules.

Main Methods:

• Developed an Instance Mask Alignment (IMA) knowledge distillation framework.
• Introduced instance mask distillation to incorporate mask information for better region focus.
• Implemented a cascade alignment module with instance standardization and adaptive scale deflation.

Main Results:

• Achieved substantial performance gains in object detection across various detector types.
• Demonstrated the effectiveness of IMA on MS-COCO, PASCAL VOC, and Cityscapes benchmarks.
• Showcased the framework's adaptability to heterogeneous detectors.

Conclusions:

• The proposed IMA framework significantly enhances object detection performance through effective knowledge distillation.
• IMA successfully reduces the teacher-student gap, offering a robust solution for diverse object detection scenarios.