Convolutional networks have had great success in image classification and other areas of computer vision. Recent efforts have designed deeper or wider networks to improve performance; as convolutional blocks are usually stacked together, blocks at different depths represent information at different scales. Recent models have explored `skip' connections to aggregate information across layers, but heretofore such skip connections have themselves been `shallow', projecting to a single fusion node. In this paper, we investigate new deep-across-layer architectures to aggregate the information from multiple layers. We propose novel iterative and hierarchical structures for deep layer aggregation. The former can produce deep high resolution representations from a network whose final layers have low resolution, while the latter can effectively combine scale information from all blocks. Results show that the our proposed architectures can make use of network parameters and features more efficiently without dictating convolution module structure. We also show transfer of the learned networks to semantic segmentation tasks and achieve better results than alternative networks with baseline training settings.
View on arXiv