Computer Vision News - October 2018

Last, in comparison with other methods, see also the leaderboard quoted above, but this time with quantitative results: Summary: The achievements of DeepLabV3+ are: ● A novel encoder-decoder module based on atrous convolution, which achieves improved segmentation along object boundary lines. ● It offers users choice of the resolution of extracted features, allowing trade-off between precision and runtime, not possible in other encoder- decoder models. ● It provides an efficient and better performing encoder-decoder network, by adapting the Xception model for segmentation and using depthwise separable convolution for both ASPP and the decoder module. ● When augmenting the training dataset with the JFT-300M dataset, DeepLabV3+ achieved top performance on the PASCAL VOC 2012 leaderboard. For more details, including design choice rationales and model variants, see here . 10 Computer Vision News Research Research