Efficient Pothole Segmentation Using Quantization and Pruning for ADAS

Real-time semantic segmentation is critical for Advanced Driver-Assistance Systems (ADAS), but deploying deep learning models on resource-constrained hardware remains a challenge. This work presents an efficient pipeline for pothole segmentation using the lightweight PP-LiteSeg model, evaluating and comparing two key optimization techniques: post-training quantization and network pruning. Results demonstrate that 8-bit integer (INT8) quantization is highly effective, achieving a 3.54x speedup in GPU compute time with only a minimal, well-controlled loss in segmentation accuracy. While network pruning also reduces model size and latency, it comes at a more pronounced cost to performance. In conclusion, while quantization offers a superior trade-off between speed and accuracy, network pruning remains a valuable strategy for applications where memory compression is the primary objective.