Using CNN to Predict the Transformed Spatial Distance for a Pair of Image Patches
Chin-Hung Teng, Zhen-Hao Yang, Ching-Hu Lu · 2020
Matching image feature points is a very fundamental task in computer vision. Unlike the previous approaches that were designed from the viewpoint of classification, our study instead employs convolutional neural network (CNN) to estimate the transformed spatial distance of two features in the image plane. This spatial distance can not only be used to classify a match into correct or incorrect, but it also provides additional information about the equality of the match. To evaluate the performance of the proposed CNN, we have generated a large dataset for network training and testing. The experimental results show that if the true transformed spatial distance of a pair of corresponding features is between 0 and 5 pixels, the root mean square error produced by our network is 0.97 pixels.