Distillation of neural network models for detection and description of key points of images
Authors:
A. V. Yashchenko,
A. V. Belikov,
M. V. Peterson,
A. S. Potapov
Abstract:
Image matching and classification methods, as well as synchronous location and map**, are widely used on embedded and mobile devices. Their most resource-intensive part is the detection and description of the key points of the images. And if the classical methods of detecting and describing key points can be executed in real time on mobile devices, then for modern neural network methods with the…
▽ More
Image matching and classification methods, as well as synchronous location and map**, are widely used on embedded and mobile devices. Their most resource-intensive part is the detection and description of the key points of the images. And if the classical methods of detecting and describing key points can be executed in real time on mobile devices, then for modern neural network methods with the best quality, such use is difficult. Thus, it is important to increase the speed of neural network models for the detection and description of key points. The subject of research is distillation as one of the methods for reducing neural network models. The aim of thestudy is to obtain a more compact model of detection and description of key points, as well as a description of the procedure for obtaining this model. A method for the distillation of neural networks for the task of detecting and describing key points was tested. The objective function and training parameters that provide the best results in the framework of the study are proposed. A new data set has been introduced for testing key point detection methods and a new quality indicator of the allocated key points and their corresponding local features. As a result of training in the described way, the new model, with the same number of parameters, showed greater accuracy in comparing key points than the original model. A new model with a significantly smaller number of parameters shows the accuracy of point matching close to the accuracy of the original model.
△ Less
Submitted 18 May, 2020;
originally announced June 2020.
