Fig. 1
From: Instance mask alignment for object detection knowledge distillation
Overview of the Instance Mask Alignment (IMA) framework for object detection knowledge distillation. Our distillation method follows a two-step process. Firstly, we extract instance feature maps from the teacher model. These feature maps are then used to generate an instance mask that aligns with the student features. To further enhance the alignment between the teacher and student features, we employ the Cascade Alignment technique, which includes Instance Standardization and Adaptive Scale Deflation. Finally, we calculate the total distillation loss.
