EAALSeg: an efficient all-aware long tail segmentation framework for comprehensive wild scene understanding

The comprehensive segmentation of visual semantics in wild scenes is crucial for the application of field robots. However, current real-time segmentation algorithms for wild scenes face the issues of accuracy deficiency and difficulty in tail category learning. To address the issues, we propose EAALSeg, an efficient all-aware long-tail segmentation framework for more comprehensive scene understanding. We introduce efficient all-aware attention block which captures feature correlations at local detail, global context, and cross-sample dimensions, while maintaining efficiency. We propose a new mixed long-tail learning strategy that optimizes data, loss, and training process simultaneously. Our experimental results on the publicly available datasets RUGD and RELLIS-3D show that our framework achieves the optimal balance between efficiency and precision while successfully addressing the issue of tail classes being unsegmentable.