Computer Vision News - February 2021

All modules are initialized from scratch with random weights and during training AdamW optimizer is used. As network loss, the distance between predicted and ground truth flow is employed: where and Several random augmentation methods are also applied to the inputs during training: 1) photometric augmentation (perturbation of brightness, contrast, saturation, and hue); 2) spatial augmentation (rescaling and stretching); 3) erosion of rectangular regions in with probability 0.5 to simulate occlusions. Further details of the training are summarized in the table below, including (first 3 columns) the combinations of data used for training and finetuning. 2 1 6 Research = ∑ −1 || − || 1 =1 = ( − 1) characteristics) and two convolutional GRU update blocks with 1x5 filters and 5x1 filters respectively, is shown below. FEATURE ENCODER CONTEXT ENCODER Input ( 1 , 2 ) 1 Normalization Instance Batch =0.8.