• Proposes ProC-KD, a novel cross-task knowledge distillation method that removes the label-space constraint between teacher and student networks.
• Introduces a prototype learning module to capture invariant intrinsic local object representations from the teacher network.
• Develops a task-adaptive feature augmentation module that enhances student features with generalized prototypes to improve generalization.
• Demonstrates effectiveness across various visual tasks, confirming the practicality of cross-task knowledge distillation.