China
Africa
Explore chapters and articles related to this topic
Multiview Image Matching for 3D Earth Surface Reconstruction
Published in Yuhong He, Qihao Weng, High Spatial Resolution Remote Sensing, 2018
In multisource image matching, image intensity is no longer a dominating matching feature, unless matching images can be adjusted toward similar spatial resolution and intensity distribution. Conversely, feature matching is usually more stable than intensity matching (Gruen, 2012) if images are obtained under different illumination and atmospheric conditions. A series of existing matching costs were evaluated in Hirschmuller and Scharstein (2009). The Census Transform (Zabih & Woodfill, 1994), which converts each pixel inside a moving window into a bit vector representing which neighbors are above or below the central pixel, was found to be robust against large-scale, nonstationary exposure and illumination changes. In Tack et al. (2012b) and Aguilar et al. (2014), cross-sensor optical images were matched after radiometric and geometric normalization of the multitemporal and multisensor imagery. In Sedaghat et al. (2012), multisource optical imagery was matched via a combination of techniques, including the SIFT algorithm, the Harris corner detector, the LSM approach, a robust control network construction technique, and a new target matching strategy based on distance and angle differences.
Gradient guided feature selection in stereo matching
Published in Amir Hussain, Mirjana Ivanovic, Electronics, Communications and Networks IV, 2015
Considering that Census transform has turned out to be effective for stereo correspondence (Mei et al. 2011), which is a non-parametric local descriptor (Zabih & Woodfill 1994), it is employed as the multi-scale features to set up the training set. Census transform evaluates the relative ordering of local intensity values within certain size of sub-window. The readers can refer to (Zabih & Woodfill 1994) for details. Specifically, during constructing the training set, Census transform with five widths are applied on each original image patch, which are respectively 3,6,9,12 and 15. The one with the best performance is marked with the label of the corresponding gradient image patch, where the performance is measured by disparity error with threshold 1. These gradient images together with their labels construct the training set for later regularized logistic regression based on feature selection training.
3D imaging
Published in Michael O’Byrne, Bidisha Ghosh, Franck Schoefs, Vikram Pakrashi, Image-Based Damage Assessment for Underwater Inspections, 2019
Bidisha Ghosh, Michael O’Byrne, Franck Schoefs, Vikram Pakrashi
Given two images of the same scene, the stereo matching algorithm aims to match pixels in one image with the corresponding pixels in the other image. A similarity measure is used to evaluate how closely a patch in the reference image (normally the left image) resembles a candidate patch in the other stereo image, and thus how likely the patches depict the same region in the scene. Patches may be compared on the basis of pixel intensity values, texture patterns, or using census/rank-transformed data. Common measures of similarity used for computing the matching cost include sum of squared distances (SSD), sum of absolute distances, normalized-cross correlation, and Hamming distance (Giachetti, 2000). There are advantages and drawbacks associated with each similarity measure. Correlation-based metrics are capable of tolerating variations in image brightness, making them a good choice for situations where challenging lighting conditions exist, or when the left and right cameras in a stereo system have different exposure settings. Likewise, census transform-based matching is known to be robust to radiometric distortion since global lighting differences between two images will not affect the ordering of pixels at a local level. The census transform summarizes local image structure. It is based on the relative ordering of local intensity values and not on the intensity values themselves. This aspect means that it can tolerate outliers and perform better near object boundaries compared with only using color data (Tavera-Vaca et al., 2015). However, these matching metrics often take a longer time to compute compared with simple similarity measure such as the sum of squared distances (SSD), which is the metric used in this demonstration.
Stereo dense depth tracking based on optical flow using frames and events
Published in Advanced Robotics, 2021
Antea Hadviger, Ivan Marković, Ivan Petrović
Our method is invariant to the choice of a dense frame-based disparity estimation algorithm. It can work in both stereo and monocular setup. However, since events are mostly generated on boundaries of image segments, it is desirable that the chosen frame-based disparity estimation method yields sharp and precise results in these areas, i.e. at disparity discontinuities. Otherwise, updating the disparity using events would not be effective. We used our implementation of semi-global matching (SGM) for a stereo setup presented in [22], which supports AVX2 instructions and multithreading to support real-time performance. This algorithm aggregates stereo matching costs along several linear paths as follows: where denotes the pixel location and d is the potential disparity. is the accumulated loss along direction , and is the matching cost between the referent image patch and the considered potential match. The final accumulated cost is the sum of all costs . In our implementation, we use the matching cost based on the census transform. SGM introduces parameters and as discontinuity penalties in order to obtain smooth disparity maps. is the penalty for a disparity change of 1 pixel, while penalizes any disparity change i larger than 1. This, however, consequently leads to loss of precision on disparity boundaries as the algorithm resists to change the disparity, which is unfavorable for our method. For this reason, we chose smaller values of and than it is usually recommended, resulting in better discontinuity sharpness, but at the price of less smoothness.