Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs

摘要: Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs probabilistic graphical models for addressing task pixel-level (also called "semantic segmentation"). We show that responses at final layer are not sufficiently localized accurate segmentation. is due to very invariance properties make good tasks. overcome this poor localization property deep networks by combining DCNN with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system able localize segment boundaries accuracy which beyond previous methods. Quantitatively, method sets new state-of-art PASCAL VOC-2012 semantic segmentation task, reaching 71.6% IOU test set. how these results can be obtained efficiently: Careful network re-purposing novel application 'hole' algorithm wavelet community allow dense computation neural net 8 frames per second on modern GPU.