Toward Extending Post-Training Quantization for CNN-Based Semantic Segmentation Models

초록

Post-Training Quantization (PTQ) is an efficient method for compressing neural networks without retraining, but its application to CNN-based segmentation models has not been thoroughly explored. Therefore, this paper applies three representative PTQ methods to PIDNet, a CNN-based semantic segmentation model, and quantitatively compares their performance. The experimental results show that all three methods maintain satisfactory performance under the 8/8-bit and 4/8-bit configurations, but exhibit a clear degradation in the 4/4-bit setting. These findings highlight the need for more sophisticated activation quantization methods tailored for dense prediction tasks under ultra-low bit configurations. © 2025 IEEE.

키워드