Learning Deconvolution Network for Semantic Segmentation

The paper introduces a semantic segmentation method that learns a deep deconvolution network on top of the convolutional layers adopted from the VGG 16-layer network. The deconvolution network consists of deconvolution and unpooling layers that identify pixelwise class labels and predict segmentation masks. Rather than segmenting the whole image at once, the trained network is applied to each object proposal, and the final segmentation map is assembled by combining the results from all proposals.

This proposal-wise, deconvolution-based design mitigates limitations of fully convolutional network approaches, letting the method recover detailed structures and handle objects across multiple scales more naturally. On the PASCAL VOC 2012 benchmark, the approach reaches 72.5% accuracy through an ensemble with a fully convolutional network, the best result among methods trained without the Microsoft COCO dataset.