Computer Vision News - October 2018

the figure below in the spatial data reconstruction process. In the spatial information reconstruction process, the corresponding encoding-level feature map among encoder-output feature maps is cropped and concatenated for more precise segmentation of object boundaries. DeepLabV3+ uses a structure similar to that of U-Net, improving object boundary prediction. Xception and Modified Aligned Xception The Xception network is based on the inception model, combining it with depthwise separable convolution -- in fact xception takes the concept of inception to the extreme (thus its name -- xception = extreme inception). The convolution layer only performs convolution for each channel separately (and Research 7 Research Computer Vision News