Search

Search Results

• THE METRICS FOR TRACKING ALGORITHMS EVALUATION

  А. Е. Shchelkunov, V.V. Kovalev, K.I. Morev, I.V. Sidko

  2020-07-10

  Abstract ▼

  The work is devoted to a review of existing metrics for assessing the quality of the task of tracking objects on video with various algorithms. When evaluating tracking algorithms for their subsequent com-parison, it is not enough to use one metric, and algorithms should be evaluated using a set of different independent estimates. To this end, a study was conducted of existing metrics for evaluating algorithms, the results of which are given in the article. The review involves many different approaches to evaluating algorithms. For example, approaches based on the assessment of the definition of the center of the track-ing object, which are one of the first and still popular metrics for evaluating tracking algorithms. The main disadvantages of such approaches include the difficulty of determining the true center of the object, as well as the interpretation of estimates for various sizes of the object. To eliminate these shortcomings, anew metric is introduced in the article: an unbiased (window) error in determining the center of an object, which takes into account the constant component of the error in determining the center. Other approach-es include metrics based on the analysis of the intersection over union. Also, the article considers ap-proaches based on the analysis of tracking failures, which take into account the tracking length and fail-ure rate. A new method is proposed for evaluating algorithms in case of loss of visual contact with an tracking object, taking into account the number of frames in which visual contact with the object was lost. During the study, approaches to evaluating algorithms for simultaneous tracking of several objects were considered. Integral metrics were proposed whose task is to obtain a comprehensive assessment of the tracking algorithm. For the formation of a comprehensive assessment, it is desirable to use various un-correlated metrics. Complex estimates provide the ability to compare algorithms with each other. As a comprehensive assessment, the article proposes the use of a metric combining the accuracy and robust-ness of the algorithm. As a rule, the intersection over union is used as the accuracy metric, however, for problems where the accuracy of tracking the center of the object is fundamental, the authors propose using an unbiased error in determining the center as the accuracy metric.

• ACCELERATION OF THE DIRECT PASSAGE IN THE IMPLEMENTATION OF CNN ON A LIMITED COMPUTING RESOURCE

  А.Е. Shchelkunov, V.V. Kovalev, I. V. Sidko, N. Е. Sergeev

  2022-04-21

  Abstract ▼

  The work is devoted to the optimization of the neural network architecture for its launch on
  a limited computing resource. Several optimization approaches are considered, estimates of the
  complexity and execution time of the forward pass of the neural network are given. Comparative
  estimates of the complexity of the network using different optimization approaches are given.
  The paper presents an analysis of the selected network architecture, and estimates of the computational
  complexity of individual components (modules) of the architecture are obtained. An analysis
  of possible optimization methods for each module was made. The parameters of the considered
  modules, the sizes of the input and output tensors are described. Several architectures were tested
  to optimize the feature extraction module, ResNet 50, ResNet 18, MobileNet v3 small, MobileNet
  v3 large. A comparative analysis of the computational complexity and execution time of the forward
  pass for each architecture is presented. Forward pass times were measured on Nvidia's
  Jetson AGX Xaver embedded computing device. Estimates of the execution time of the direct pass
  for each module of the considered neural networks are presented. The paper presents the results of
  comparing neural network accuracy estimates before and after architecture optimization. The test
  data set consists of 100 video recordings. 5 different typical objects are involved in test videos,
  10 different scenarios are recorded for each object class. For each of the developed architectures,
  accuracy estimates were obtained, and a comparative analysis was made.