Answers are grounded in this paper and cite their sources.
Martian surface understanding plays a crucial role in planetary geoscience, supporting autonomous rover navigation and geological interpretation in deep-space exploration. However, existing deep learning methods for Martian surface analysis are typically designed for single tasks and fail to exploit shared representations across terrain detection and segmentation. Although a few multitask frameworks incorporate auxiliary geometric cues, they lack the capability for unified instance-level recognition and fail to handle the extreme scale variations and heterogeneous task imbalances inherent to Martian terrains. In this study, we propose MarsMTFormer, a unified multitask transformer framework for Martian surface object detection and instance segmentation. To address the large-scale variations and sparse contextual information in Martian terrains, we design a gated global context enhancement decoder, which introduces compact global semantic tokens and employs an adaptive gating mechanism to selectively inject scene-level context into query representations, resulting in more stable and discriminative features. Moreover, the severe long-tailed distributions of Mars data lead to heterogeneous imbalance across detection and segmentation tasks, posing significant challenges for unified multitask learning. To overcome this, we develop an adaptive multitask weighted sampling strategy that dynamically balances training by jointly modeling category frequency and task-specific performance feedback, enabling coordinated optimization across detection and segmentation tasks. Extensive experiments demonstrate that MarsMTFormer delivers competitive performance on the MarsScapes and SynMars-TW datasets.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: