Opt-RetinaSeg: Optimized Semantic Segmentation for Embedded Automotive Systems
Opt-RetinaSeg is an innovative architecture that modifies the RetinaNet detection framework to enable real-time semantic segmentation on embedded automotive systems. It substitutes ResNet-50 with a hybrid, lightweight feature extractor, reconfigures the Feature Pyramid Network to minimize unnecessary computations, and features a compact segmentation head that employs focal-loss-inspired class balancing specifically for road environments. Additionally, a three-stage optimization process implements structured channel pruning along with post-training INT8 quantization. This research can be found on arXiv with the ID 2607.22714.
Key facts
- Opt-RetinaSeg is derived from RetinaNet for dense pixel-wise prediction.
- The backbone is changed from ResNet-50 to a hybrid lightweight feature extractor.
- Feature Pyramid Network is restructured to reduce multi-scale computation.
- Segmentation head uses focal-loss-inspired class balancing for foreground-background imbalance.
- Three-stage optimization includes structured channel pruning and INT8 quantization.
- Targeted at embedded automotive systems with compute, memory, and power constraints.
- Paper ID: arXiv:2607.22714.
- Published on arXiv.
Entities
Institutions
- arXiv