I present novel supervised and semi-supervised semantic segmentation algorithms by learning deconvolution networks, which are recently investigated in POSTECH Computer Vision Lab. The deconvolution network is constructed on top of the convolutional layers adopted from VGG 16-layer net. It is composed of deconvolution and unpooling layers, which identify pixel-wise class labels and predict segmentation masks. Our supervised algorithm mitigates the limitations of the existing methods based on fully convolutional networks by integrating deep deconvolution network and proposal-wise prediction; our segmentation method typically identifies detailed structures and handles objects in multiple scales naturally. Also, our semi-supervised technique decouples classification and segmentation procedure, which facilitates to employ rich information of weak image-level annotations and achieve very nice performance even with very small number of training examples with strong annotations. Our networks demonstrate outstanding performance in PASCAL VOC 2012 dataset, especially in semi-supervised setting.