Feature Fusion Use Unsupervised Prior Knowledge to Let Small Object Represent

TL;DR

This paper introduces FillIn, a novel feature fusion method leveraging superpixel-based prior knowledge to enhance small object representation in segmentation tasks.

Contribution

The paper proposes a new fusion mechanism called FillIn that uses superpixel segmentation to improve small object segmentation by aligning low and high level features.

Findings

Achieves competitive results on PASCAL VOC 2012

Enables small objects to be better represented in segmentation

Visualizations confirm improved small object segmentation

Abstract

Fusing low level and high level features is a widely used strategy to provide details that might be missing during convolution and pooling. Different from previous works, we propose a new fusion mechanism called FillIn which takes advantage of prior knowledge described with superpixel segmentation. According to the prior knowledge, the FillIn chooses small region on low level feature map to fill into high level feature map. By using the proposed fusion mechanism, the low level features have equal channels for some tiny region as high level features, which makes the low level features have relatively independent power to decide final semantic label. We demonstrate the effectiveness of our model on PASCAL VOC 2012, it achieves competitive test result based on DeepLabv3+ backbone and visualizations of predictions prove our fusion can let small objects represent and low level features have potential for segmenting small objects.