Convolutional neural networks (CNNs) have recently been used for a variety of computer vision tasks,including pose regression and image classification. Estimating the absolute localisation of a monocular camera in six degrees of freedom (6-DOF) is a new CNN application. The accuracy of the data set utilized will directly affect how well this strategy performs. In this research, utilizing a CNN architecture built on six dimensions of pose regression, we present new metrics to assess dataset quality and achieve adequate performances of Camera absolute posture estimate from a single image. The Structural SIMilarity (SSIM) and Correlation similarity rates are the suggested metrics. These two metrics are based on the analysis of visual similarity across all of the CNN training images. By leveraging the camera poses in our work, we present a different type of metric. As a result, using the position and orientation of the camera, all training camera views are combined to create a three-dimensional map of a probable coverage volume. The camera covering rate metric is used to define the latter.The proposed metrics are intended to give a first understanding of a dataset’s limitations when utilized for absolute pose regression using CNN architecture. By assuring an acceptable Camera coverage volume and minimizing ambiguity by managing similarity carefully, our metrics may also direct us effectively in dataset elaboration to cope with the expected performances. Our metrics are evaluated using the well-known PoseNet framework, which is applied to a well-known dataset (7-scenes).The obtained results are encouraging and deal with our expectations in terms of prediction errors.
MSC Classification: 68T07 , 68T01 , 68T20 , 68T45