A Comparison of State-of-the-Art Network Architectures for Instance-Segmentation in Forest Environments

Research and development have increasingly focused on automating mobile machines to reduce the negative influence of labor shortages and high labor costs. Object detection is a key requirement for the automation of mobile machines. The transfer of the developed methods to the environment of mobile machines, e.g. a forest, a building site, or in mining, is challenging. Objects of the same class can have significantly different phenotypes and the surroundings cannot be controlled, weather as well as lighting conditions can change. Neural networks are the state-of-the-art method for detecting and classifying objects for image sensors. The required datasets as well as network architectures mastering object detection across different forest areas have not yet been presented. We collected two datasets, MobimaWoodlands and MobimaSkidRoads, one with a handheld camera and one captured while driving on skid roads in different areas and in different seasons. Three network architectures for the instance segmentation with two different backbones were trained on the two datasets to segment stems, trees, and stumps. In a subsequent step, the trained networks were evaluated on two public datasets which have not been used in the training process. ... mehrWith an adapted training pipeline, we achieved a similar accuracy with a slight decrease in the AP of 0.1 on one of the unknown datasets with similar tree specimens. On the second unknown dataset, the AP decreased more significantly by 0.3. The findings highlight that generalization over various forest areas is possible, even in demanding outdoor settings. However, the portability to unknown domains cannot be guaranteed especially if different tree species are present, which continues to be an issue in many applications.