Feature Detection Performance Based Benchmarking of Motion Deblurring Methods: Applications to Vision for Legged Robots

Multi‐Frame motion deblurring of video using the natural oscillatory motion of dexterous legged robots

“…Their method can achieve better running times compared to maximum a posteriori estimation based methods. In a prior study [2], we have compared most of these methods in both controlled motion scenarios as well as in realistic robot motion scenarios under identical conditions. We have selected the leading methods for our comparative results in the present study.…”

“…This data is the basis of our experiments which are presented in Section 4. The details of the dataset, which is available from the authors for research use, can be found in [2]. All of the experiments below are conducted using the aforementioned dataset with off-line computation.…”

“…With this controlled 1-D linear motion setup, it is possible to obtain pairs of differently blurred frames that can be used in the multi-frame deblurring experiments with different motion blur cases. The tests are carried out using a dataset [2] having 20, 30, 50, 60, and 70 pixels blur magnitudes with 15 different frames from each blur magnitude. The results are compared using the FD-AROC metric [2] which indicates the quality of a deblurred image in terms of correctly detecting the image features utilising the receiver operating characteristic (ROC) curve.…”

“…This is opposed to kinematically stable legged robots that are much slower and maintain static stability through carefully selected leg positioning. The agile locomotion of dexterous legged platforms causes oscillations of the robot body which may become more severe depending on the surface properties and different locomotion speeds [2]. During robot operation in a given environment, there is ever present need to use body-mounted sensors to collect information about the robot egostate as well as features of the environment.…”

“…1b and involves both distortion and loss of information in the captured image data. Motion blur is a highly undesirable phenomenon for computer vision applications that severely degrades the performance of almost all feature extraction and perception algorithms [2].…”