Rozdziały
Stały URI dla kolekcjihttp://hdl.handle.net/11652/4775
Przeglądaj
4 wyniki
collection.search.results.head
Pozycja On the Importance of the RGB-D Sensor Model in the CNN-based Robotic Perception(Wydawnictwo Politechniki Łódzkiej, 2023) Zieliński, Mikołaj; Belter, DominikMobile and manipulation robots operating indoors use RGB-D cameras as the environment perception sensors. To process data from RGB and depth cameras neural networks are applied. These neural-based systems are trained using synthetic datasets due to the difficulties of obtaining ground truth data on real robots. As a result, the neural model used on the real robot does not produce satisfactory performance due to the differences between the images used during training and the inference. In this paper, we show the importance of depth sensor modeling while training the neural network on a synthetic dataset. We show that the obtained neural model can be used on the real robot and process the data from the real RGB-D camera.Pozycja NeRF-based RGB-D Images Generation in Robotics – Experimental Study(Wydawnictwo Politechniki Łódzkiej, 2023) Kulecki, Bartłomiej; Belter, DominikMultiple learning-based algorithms in robotics require collecting RGB-D images of the scene from various viewpoints. These procedures are time-consuming, so many methods are trained using synthetic images. Recently, a Neural Radiance Fields (NeRF) model of the scene was proposed. Moreover, recent methods show that this model can be trained in minutes. This opens the possible applications in robotics for training the systems to reconstruct scenes, grasp objects or estimate their 3D poses using RGB-D images generated from a small number of input images. In this paper, we verify the quality of RGB-D images generated by the Instant Neural Graphics Primitives implementation of NeRF. We compare the obtained results from the Instant NeRF with the ground-truth RGB-D images obtained from the Kinect Azure and images generated from the point cloud model of the scene. The results show that the difference between generated RGB-D images and ground truth images is small, especially near the object.Pozycja Mixing Synthetic and Real-world Datasets Strategy for Improved Generalization of the CNN(Wydawnictwo Politechniki Łódzkiej, 2023) Młodzikowski, Kamil; Belter, DominikIn this paper, we deal with the problem of supervised training neural networks with an insufficient number of real-world training examples. We propose a method that at the beginning trains the neural network using a relatively simple synthetic dataset. In the following epochs, we add more challenging and real-life images to the training dataset. We compare the proposed strategy with other methods of using artificial and real-world datasets for training the neural network. The obtained results show that the proposed strategy allows for obtaining the neural network with higher generalization capabilities than competitive methods.Pozycja 3D Reconstruction of Non-Visible Surfaces of Objects from a Single Depth View – Comparative Study(Wydawnictwo Politechniki Łódzkiej, 2023) Staszak, Rafał; Michałek, Piotr; Chudziński, Jakub; Kopicki, Marek; Belter, DominikScene and object reconstruction is an important problem in robotics, in particular in planning collision-free trajectories or in object manipulation. This paper compares two strategies for the reconstruction of nonvisible parts of the object surface from a single RGB-D camera view. The first method, named DeepSDF predicts the Signed Distance Transform to the object surface for a given point in 3D space. The second method, named MirrorNet reconstructs the occluded objects’ parts by generating images from the other side of the observed object. Experiments performed with objects from the ShapeNet dataset, show that the view-dependent MirrorNet is faster and has smaller reconstruction errors in most categories.