Real-Time Semantic Segmentation with Optimized RetinaNet Architectures for Embedded Automotive Systems

This paper presents Opt-RetinaSeg, a semantic-segmentation architecture derived from RetinaNet for resource-constrained automotive hardware. It replaces the ResNet-50 backbone with a hybrid lightweight extractor, simplifies the feature pyramid network, and uses a compact segmentation head with focal-loss-inspired class balancing. The reported optimization pipeline combines structured channel pruning, post-training INT8 quantization, and knowledge distillation. On Cityscapes and BDD100K, the authors report 73.9% mIoU at 70.4 FPS, a 7.4x inference speedup, and a 4x model-size reduction versus a ResNet-50 baseline, with less than 3% accuracy degradation, deployed on Jetson Xavier NX and Qualcomm QCS610 hardware.