Multi-scale Discriminative Location-Aware Network for Few-Shot Semantic Segmentation

Few-shot semantic segmentation methods aim for predicting the regions of different object categories with only a few labeled samples. It is difficult to produce segmentation results with high accuracy when a new category appears. In this paper, we propose a Multi-scale Discriminative Location-aware (MDL) network to tackle the few-shot semantic segmentation problem. In order to use information from different levels, we first keep the last three convolutional layers of FCN, and then use the VGG-16 network to extract features from the support image-label pair, which adjusts the weight of the query image segmentation branch. Discriminative location-aware architecture can improve the efficiency of few-shot segmentation, and therefore the global average pooling layer is added to produce location feature information. Finally, we evaluate our MDL model on the Pascal VOC 2012 challenge, and show that it achieves competitive mIoU score compared to methods in recent years.

I. Introduction

Models based on convolutional neural networks have become popular for boundary detection and semantic segmentation problems in computer vision. However, one of their main shortcomings is that they need a large number of manually annotated images. These annotated approaches such as semantic labeling and instance segmentation cost tremendous manpower and have low effectiveness. To reduce the budget, bounding box-level and image-level labeling methods are proposed to make the annotation task cheaper and easier than pixel-level mask. However, existing techniques for training the label from bounding boxes predict segmentation results with high noise, and they lack the general ability to predict unknown categories. Figure 1 shows the pixel-level segmentation mask and rectangle mask, and figure 1(b) can be regarded as a particular kind of label for weakly-supervised learning.