This is the authors' abstract. We don't add key points for this paper.
Semantic Segmentation has been intensely studied for scene understanding in case of structured environments like highways and city streets. However, when it comes to unstructured environments such as forests, grasslands, agricultural lands, etc. the 3D scene becomes more challenging with lack of conclusive object boundaries, shadow occlusions, overlapping colors or textures of different obstacles, etc. This makes the process of semantic segmentation more complex in these scenarios. The traditional segmentation models make use of Convolutional Neural Networks (CNN) for feature extraction. However, for more scene understanding in off road terrains with complicated backgrounds and uncertain obstacles, the model needs to be able to have an overall comprehension of the entire scene in addition to the local pixel clusters. Our model uses transformer-based fusion module in the backbone to fuse extracted semantic and geometric features from images and point clouds that enable contextual understanding making complex semantic segmentation more accurate. Moreover, multi-resolution fusion of features from camera and LiDAR increases the reliability of our perception system by ensuring consistent information availability at all times. Our results on the GOOSE offroad dataset indicate that our model performs better than other state-of-the-art segmentation models in accuracy by 2% and precision by 1.58%.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Kunjumon et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: