ESANet: Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis

Our network architecture for efficient scene analysis ESANet enables real-time semantic segmentation with up to 29.7 FPS on Jetson AGX Xavier. [...] ESANet achieves a mean intersection over union of 50.30 and 48.17 on [indoor datasets NYUv2 and SUNRGB-D]. Our models are trained with PyTorch, [...] exported to ONNX [and] converted to TensorRT engines. During network design, we [...] only use operations [...] supported and highly optimized by TensorRT, [enabling] up to 5× faster inference compared to pure PyTorch. ESANet is well suited as a common initial processing step in a complex system for real-time scene analysis on mobile robots.