General Saliency Detection Model for 2D and Panoramic Images
Ripei Zhang, Chunyi Chen, Jun Peng · 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA) · 2022
Panoramic images have geometric distortions different from 2D images, so the current significance detection model cannot be used in two forms of images. Although viewers need to use head-worn devices to watch panoramic images, human visual attention mechanism will not change with the change of observation mode. Therefore, we propose a significance detection model that shares weights between two images, so that it can be applied to 2D images and panoramic images at the same time. Our model uses different convolution sampling modes for 2D images and panoramic images respectively, so that the saliency detection model based on the same set of parameters can be correctly applied to the images of two modes at the same time. Firstly, we separate the sampling and calculation steps in spherical mapping convolution, and then design two different sampling methods to make them suitable for 2D and panoramic images respectively. Finally, we use 2D convolution to convolute the sampled results. In addition, in order to eliminate the influence of 2D and panoramic data set annotation methods on the calculation results, we annotate the two kinds of data sets in a unified way. The experimental results show that our model can be applied to 2D and panoramic images, and the detection accuracy is better than the existing models under the condition of less training times.