A hybrid attention semantic segmentation network for unstructured terrain on Mars

Semantic segmentation of Martian terrain is crucial for the route planning and autonomous navigation of rovers on Mars. However, existing methods are restricted to structured or semi-structured scenes, performing poorly on Mars that is a completely unstructured environment. Therefore, we propose a novel hybrid attention semantic segmentation (HASS) network, which contains a global intra-class attention branch, a local inter-class attention branch and a representation merging module. Specifically, the global attention branch draws the consistencies of all homogeneous pixels in the whole image, and the local attention branch models the relationships between specific heterogeneous pixels with the supervision of elaborately designed loss function. The merging module aggregates the contexts from the two branches for the final segmentation. Furthermore, we establish a panorama semantic segmentation dataset of Martian landforms, named MarsScapes, which provides fine-grained annotations for eight semantic categories. Extensive experiments on our MarsScapes and the public AI4Mars datasets show the superiority of the proposed method.

Automated summary

Semantic segmentation of Martian terrain is crucial for rover route planning and navigation on Mars. Existing methods perform poorly in the unstructured Martian environment. Therefore, a novel hybrid attention semantic segmentation (HASS) network is proposed, which includes a global intra-class attention branch, a local inter-class attention branch, and a representation merging module. In addition, a panorama semantic segmentation dataset named MarsScapes is established, providing fine-grained annotations for eight semantic categories. Extensive experiments on MarsScapes and the public AI4Mars datasets demonstrate the superiority of the proposed method.