Semantic analysis is a novel and fast-growing field of computer vision which automates the extraction of image and scene descriptions according to human perception. This is essential to broaden the uses of computer vision to human-like tasks, providing more meaningful descriptions than the traditional low-level properties and features of the scene. Semantic segmentation is a crucial component of semantic analysis as it predicts class labels for each individual sensory data point, providing a rich analysis of scene semantics.
From a paper by Kelian J.L. Massa et al.
Researchers have made significant progress in solving the problem of semantic segmentation in both two-dimensional (2D) images and three-dimensional (3D) scenes. However, most of the work on semantic segmentation of 3D scenes has been performed on artificial and urban scenes [1], containing large amounts of man-made structures. Off-road scenes with large amounts of natural scenery and vegetation have seen far less usage in comparison. Urban scenery tends to contain distinct objects with structured features, clearly defined boundaries, and relatively even distributions of different classes. While most datasets contain some vegetation and natural scenery, these are often grouped into a few broad categories such as terrain and vegetation, making them easy to distinguish from the surrounding urban scenery.
On the other hand, natural scenes typically exhibit a degree of randomness and fluidity, marked by an irregular distribution of classes where a few classes dominate the scene’s overall composition. An off-road setting, such as a natural reserve or farm, often consists of diverse and uneven terrain along with a wide variety of objects, typically in limited quantities. This results in a markedly imbalanced distribution of class labels. Coupled with the inherent variability and irregularity of natural objects, achieving high intersection over-union scores becomes notably more demanding. Consequently, the problem of 3D semantic segmentation of natural scenes is yet to be solved with reasonable accuracy [2], [3].
Despite the inherent complexities, humans are able to easily perceive natural environments. Expanding the work done in 3D semantic segmentation specifically to natural scenes would provide for significantly more flexible and adaptable computer vision based systems. This would largely have applications in robotics that requires some interpretation of surroundings in natural environments. For instance, autonomous vehicles such as self-driving cars operating in off-road environments, autonomous tractors used in farming, and surveillance drones operating in natural environments, could benefit from the ability to interpret and understand their surroundings more accurately [4], [5].
For the full paper CLICK HERE
Note – If you liked this post click here to stay informed of all of the 3D laser scanning, geomatics, UAS, autonomous vehicle, Lidar News and more. If you have an informative 3D video that you would like us to promote, please forward to editor@lidarnews.com and if you would like to join the Younger Geospatial Professional movement click here