ViRFD: a virtual-realistic fused dataset for rock size analysis in TBM construction

In TBM (Tunnel Boring Machine) construction process, the rock size analysis system plays an important role in assisting driving. Its core algorithm is based on semantic segmentation, and it brings challenges to dataset acquisition in real applications. To relieve this problem, this paper proposes a virtual-realistic fused dataset, short for ViRFD. The R-part is composed of a realistic dataset from our previous work, and the V-part is simulated by a learning-based method proposed in this paper. Unlike traditional manual methods, we use a virtual engine (Unity) to simulate datasets, since the corresponding ground-truth labels can be automatically extracted by the engine. Specifically, we propose a novel synthetic dataset simulator, named RockSegX . It contains abundant virtual 3D resources to ensure the diversity and fidelity of generated datasets. The main feature of RockSegX lies in its content flexibility, i.e., we are able to control the content of dataset by adjusting the values of several attributes. These attributes are carefully designed for reducing the content difference between V-part and R-part datasets. And we employ a learning-based method to automatically adjust the attributes so that the V-part dataset has the smallest content difference with the R-part. Experimental results show the effectiveness of our method in improving the quality of simulated dataset, and it further boosts the test accuracy for real-world segmentation.