The solutions obtained by training the deep neural network are highly dependent on the parameters including the learning rate. Therefore, finding the appropriate settings for training deep neural networks is very important. In particular, it is necessary to find the better settings for SOTA models of Vision Transformer(ViT), whose structure is different from ordinal models. In this paper, we focus on the learning rate to find a better value using the Learning Rate Range Test (LRRT). Through our experiments, we found that the appropriate LR is located where the decrease in loss value stops in the LRRT. In addition, we discuss about the effects of the number of epochs and the LR warm up.
Image segmentation is one of the most important techniques in computer vision and image processing. Many image segmentation methods have been proposed for these few decades. Hierarchical Feature Selection (HFS)1 is a graph-based approach for the image segmentation. It is known as a fast segmentation method that merges over segmented regions hierarchically. At the first level of the merge, the superpixels are utilized to obtain the over segmented regions. However, HFS sometimes fails when it is applied for the textured regions. In this paper, we propose a new approach for image segmentation, Searching Tree Segmentation from Superpixel (STSS), by formulating the merge of superpixels as a path searching problem. We construct trees and search the trees whose nodes correspond to the boundary of the superpixels and values of the nodes correspond to the distance between superpixels. Our algorithm does not check the boundaries of similar superpixels if these are no neighboring boundaries of the distinctively different superpixels to prevent the over segmentation of the textured regions, while HFS checks all boundaries including quite similar superpixels.
Antique stereographs were taken by photographers more than 100 years ago. In general, the antique stereographs are printed to the about 9cm×17cm thick paper and they are distorted due to aged paper surface. To archiving them, their captured images have to be transformed to the original image. In the paper, a user support system for archiving antique stereographs by distorting input image of a stereograph. In our system, a user selects by clicking multiple points at the edge of the image. The input image is divided into small regions based on the points and each small region is transformed to the same rectangle with the known size by a perspective transformation. To obtain small regions all of which are originally the rectangle with the same size, the points obtained by user’s click are interpolated quadratically. To evaluate the correction accuracy of the proposed system, the experimental results are shown.
Range image registration is an essential technique for 3D modeling of the real world object from its range images captured by a 3D sensor. It estimates the relative positions and orientations of the viewpoints using multiple range images which are captured by the 3D sensor from multiple viewpoints. The rigid transformations to define the relative positions and orientations of the viewpoints are estimated based on the corresponding 3D points, which observe the same point on the object surface. In this paper, we propose a method for range image registration based on the consistency of the rigid transformation between corresponding point pairs to improve accuracy. We generate triplets of corresponding point pairs and evaluate the consistency because at least three corresponding point pairs are required to estimate a rigid transformation. We estimate the rigid transformation using each triplet of the corresponding point pairs and evaluate the distance between estimated rigid transformations in the parameter space. We select the triplet which have the largest number of consistent triplets and estimate the rigid transformation using the selected triplet and its consistent triplets. In the experiment, our method obtained the most accurate and the most stable registration results compared with the conventional methods, Iterative Closest Point and Globally Optimal ICP.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.