We are happy to share our recent work on efficient RGB-D semantic segmentation.
Our carefully designed network architecture for efficient scene analysis – ESANet – enables real-time semantic segmentation with up to 29.7 FPS on a NVIDIA Jetson AGX Xavier.
On the common indoor datasets NYUv2 and SUNRGB-D, our ESANet achieves a mean intersection over union of 50.30 and 48.17, respectively.
Our models are trained with PyTorch and afterwards exported to ONNX. The resulting ONNX models are converted to TensorRT engines. During network design, we paid attention to only use operations that are supported and highly optimized by TensorRT. This enables up to 5× faster inference compared to pure PyTorch.
Due to the fast runtime and robust semantic segmentation, our ESANet is well suited as a common initial processing step in a complex system for real-time scene analysis on mobile robots:
Our approach can also be applied to outdoor scenarios such as Cityscapes:
Analyzing scenes thoroughly is crucial for mobile robots acting in different environments. Semantic segmentation can enhance various subsequent tasks, such as (semantically assisted) person perception, (semantic) free space detection, (semantic) mapping, and (semantic) navigation. In our paper, we propose an efficient and robust RGB-D segmentation approach that can be optimized to a high degree using NVIDIA TensorRT and, thus, is well suited as a common initial processing step in a complex system for scene analysis on mobile robots. We show that RGB-D segmentation is superior to processing RGB images solely and that it can still be performed in real time if the network architecture is carefully designed. We evaluate our proposed Efficient Scene Analysis Network (ESANet) on the common indoor datasets NYUv2 and SUNRGB-D and show that we reach state-of-the-art performance while enabling faster inference. Furthermore, our evaluation on the outdoor dataset Cityscapes shows that our approach is suitable for other areas of application as well. Finally, instead of presenting benchmark results only, we also show qualitative results in one of our indoor application scenarios.
Seichter, D., Köhler, M., Lewandowski, B., Wengefeld T., Gross, H.-M. Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis arXiv preprint arXiv:2011.06961 (2020).
Note that the preprint was accepted to be published in IEEE International Conference on Robotics and Automation (ICRA).
Our code as well as the network weights are available at: