This paper studies the effect of integrating color equivariance and invariance into object detection, in particular into the Faster R-CNN architecture. To better understand the influence of this integration, we introduce modifications to the traditional convolutional layers of the standard Faster R-CNN model. By employing group theory in a similar way as Group Equivariant Convolutional Networks (G-CNNs), we replace the convolution operations with operations that are equivariant to hue transformations. The modified models are tested on several different datasets in which variations and imbalances in color distributions are present. Our toy experiments demonstrate that the replacement of the convolutional layers can lead to significant improvements in performance, especially in scenarios where the data contains a substantial amount of color variation. The findings of this work suggest that incorporating color equivariance and invariance into the design of convolutional layers can enhance object detection, proposing interesting possibilities for future research on real-world tasks.

With 3D reconstruction from borescope videos of jet engines, inspections can be made more efficient. Being able to reconstruct 3D models of the jet engines potentially contributes to making the inspections autonomous. Methods that will be used for this reconstruction require to be evaluated to ensure the accuracy of the 3D models that they create. Therefore, this study investigates how a 3D reconstruction method can be quantitatively evaluated using ground truth. From the results can be concluded that 3D models that represent ground truth data can be generated using a combination of manual and algorithmic feature matching between the frames of borescope videos. Furthermore, the Wasserstein distance is found to potentially be a viable measure for quantifying the comparison between 3D models, which is needed for assessment.