Image Processing Techniques to Improve Deep 6DoF Detection in RGB Images

Abstract

Six degrees of freedom (6DoF) Object Detection has great relevance in computer vision due to its use in applications on several areas, such as augmented reality and robotics. Even with the improved results provided by deep learning techniques, object detection of textured and non-textured objects is still a challenge. The objective of this work was to seek improvements in the six degrees of freedom detection of non-textured objects using a Convolutional Neural Network (CNN) approach through the preprocessing of the images that were used for training the network. A State of the art research was carried out on techniques that use CNN to detect objects in six degrees of freedom. We also searched for filters with enhancement factors for detection. Finally, a detection technique based on a CNN was selected and adapted to use single-channel images (grayscale) as input, instead of using three-channel images (RGB) as in the original proposition, aiming to increase its robustness while reducing the complexity of the input images. The technique was also tested with the application of two different preprocessing filters to enhance the objects’ contours on the single-channel images, one being the ”pencil effect”, and the other based on local binary patterns (LBP). With this study, it was possible to evaluate the impact on the CNN detection performance due to the application of both of the filters. The proposed technique used with one channel images and the filters on the images still could not surpass the results of the technique with the three-channel image (RGB), although it indicated paths for improvement. The pencil filter also proved to be more robust than the LBP filter, as expected.