Efficient semantic segmentation via logit-guided feature distillation

Abstract

Knowledge Distillation (KD) is a critical technique for model compression, facilitating the transfer of implicit knowledge from a teacher model to a more compact, deployable student model. KD can be generally divided into two categories: logit distillation and feature distillation. Feature distillation has been predominant in achieving state-of-the-art (SOTA) performance, but recent advances in logit distillation have begun to narrow the gap. We propose a Logit-guided Feature Distillation (LFD) framework that combines the strengths of both logit and feature distillation to enhance the efficacy of knowledge transfer, particularly leveraging the rich classification information inherent in logits for semantic segmentation tasks. Furthermore, it is observed that Deep Neural Networks (DNNs) only manifest task-relevant characteristics at sufficient depths, which may be a limiting factor in achieving higher accuracy. In this work, we introduce a collaborative distillation method that preemptively focuses on critical pixels and categories in the early stage. We employ logits from deep layers to generate fine-grained spatial masks that are directly conveyed to the feature distillation stage, thereby inducing spatial gradient disparities. Additionally, we generate class masks that dynamically modulate the weights of shallow auxiliary heads, ensuring that class-relevant features can be calibrated by the primary head. A novel shared auxiliary head distillation approach is also presented. Experiments on the Cityscapes, Pascal VOC, and CamVid datasets show that the proposed method achieves competitive performance while maintaining low memory usage. Our codes will be released in https://github.com/fate2715/LFD.