Dense High-Resolution Siamese Network for Weakly-Supervised Change Detection

Traditional change detection (CD) often requires manual labeling of pixel-level data, which is time consuming and expensive. In this paper, we present a weakly supervised approach that only requires image-level labels to detect and segment changes in a pair of images. We first employ a Siamese network to generate high-resolution feature maps. Inspired by the Seed, Expand and Constrain principles [1], we introduce the weighted global average pooling (WGAP) and a one-class cross entropy loss, to produce localization maps from image-level supervision. Then we construct a fully-connected conditional random field (CRF) model to constrain the boundaries of segmentations. Our evaluations on CDnet-2014 datasets demonstrate superior detection and localization performance.