Towards Comparable Knowledge Distillation in Semantic Image Segmentation

TL;DR

This paper analyzes the challenges in comparing knowledge distillation methods for semantic image segmentation due to inconsistent training setups and hyperparameter tuning, and proposes a standardized baseline for fair evaluation.

Contribution

It identifies issues in current comparisons, highlights the importance of hyperparameter tuning, and establishes a robust baseline for future research in the field.

Findings

Hyperparameter tuning significantly affects distillation performance.

Most existing techniques do not outperform the proposed baseline.

Standardized evaluation improves comparability across studies.

Abstract

Knowledge Distillation (KD) is one proposed solution to large model sizes and slow inference speed in semantic segmentation. In our research we identify 25 proposed distillation loss terms from 14 publications in the last 4 years. Unfortunately, a comparison of terms based on published results is often impossible, because of differences in training configurations. A good illustration of this problem is the comparison of two publications from 2022. Using the same models and dataset, Structural and Statistical Texture Distillation (SSTKD) reports an increase of student mIoU of 4.54 and a final performance of 29.19, while Adaptive Perspective Distillation (APD) only improves student performance by 2.06 percentage points, but achieves a final performance of 39.25. The reason for such extreme differences is often a suboptimal choice of hyperparameters and a resulting underperformance of the student model used as reference point. In our work, we reveal problems of insufficient hyperparameter tuning by showing that distillation improvements of two widely accepted frameworks, SKD and IFVD, vanish when hyperparameters are optimized sufficiently. To improve comparability of future research in the field, we establish a solid baseline for three datasets and two student models and provide extensive information on hyperparameter tuning. We find that only two out of eight techniques can compete with our simple baseline on the ADE20K dataset.