A Ground Truth-Free Metric for Evaluating Depth Estimation Techniques

Although monocular depth estimation is the mainstream technique for depth estimation, existing methods for its evaluation are still limited. This is because it is hard to get ground-truth depth maps except for specific datasets with a small number of samples. This paper proposes a novel deep learning-based technique for evaluating depth estimation models without ground-truth. The basic idea is to use object detection and segmentation model to analyze specific characteristics of the depth maps. In particular, we address edges preservation for the target objects. Hence, we can perform model evaluation for objects across the different depth maps without the existence of ground-truth. We validated the proposed metric by showing the accuracy of four different depth estimation models using ground truth depth maps from KITTI benchmark dataset. Experimental results present the effectiveness of the proposed evaluation metric which highlights the promise of our approach for depth maps evaluation.