Incorporating multi-scale features to deep convolutional neural networks (DCNNs) has been a key element to achieve state-of-art performance on semantic image segmentation benchmarks. One way to extract multi-scale features is by feeding several resized input images to a shared deep network and then merge the resulting multi-scale features for pixel-wise classification. In this work, we adapt a state-of-art semantic image segmentation model with multi-scale input images. We jointly train the network and an attention model which learns to softly weight the multi-scale features, and show that it outperforms average- or max-pooling over scales. The proposed attention model allows us to diagnostically visualize the importance of features at different positions and scales. Moreover, we show that adding extra supervision to the output of DCNN for each scale is essential to achieve excellent performance when merging multi-scale features. We demonstrate the effectiveness of our model with exhaustive experiments on three challenging datasets, including PASCAL-Person-Part, PASCAL VOC 2012 and a subset of MS-COCO 2014.

本论文研究了怎样在卷积神经网络中实现多尺度特征表示，提出了一种注重对每个像素位置进行加权的注意力机制。通过对三个大型数据集的广泛实验，证明了该方法的有效性。

关注尺度：尺度感知的语义图像分割