Automatization of 3D Indoor Data Annotation

There are few datasets with annotation for training and benchmarking of neural network algorithms for 3D object recognition. The reason for this is the laboriousness of manual labeling of 3D bounding boxes. In this regard, methods for automatic labeling of missing modalities are becoming more and more in demand. As a rule, auto-completion methods use the modalities that exist in the datasets. This paper proposes a new method for estimating 3D bounding boxes based on the 2D annotation often found in datasets. The problem of automatic annotation in this work is formulated as a global optimization problem. The optimized 3D bounding boxes are projected and compared to the ground truth 2D bounding boxes. The optimization step is based on the difference between the projections and the ground truth 2D bounding boxes. For better convergence of the optimization process, a new loss function is implemented. In addition to developing the method, experiments were conducted on synthetic data and ScanNet dataset to demonstrate how much marked-up 2D data is needed to obtain acceptable 3D annotation quality.