Abstract
Training deep learning-based change detection (CD) model heavily depends on labeled data. Contemporary transfer learning-based methods to alleviate the CD label insufficiency mainly upon ImageNet pre-training. A recent trend is using remote sensing (RS) data to obtain in-domain representations via supervised or self-supervised learning (SSL). Here, different from traditional supervised pre-training that learns the mapping from image to label, we leverage semantic supervision in a contrastive manner. There are typically multiple objects of interest (e.g., buildings) distributed in varying locations in RS images. We propose dense semantic-aware pre-training for RS image CD via sampling multiple class-balanced points. Instead of manipulating image-level representations that lack spatial information, we constrain pixel-level cross-view consistency and cross-semantic discrimination to learn spatially-sensitive features, thus benefiting downstream dense CD. Apart from learning illumination invariant features, we fulfill consistent foreground features insensitive to irrelevant background changes via a synthetic view using background swapping. We additionally achieve discriminative representations to distinguish foreground land-covers and others. We collect large-scale image-mask pairs freely available in the RS community for pre-training. Extensive experiments on three CD datasets verify the effectiveness of our method. Ours significantly outperforms ImageNet, in-domain supervision, and several SSL methods. Empirical results indicate ours well alleviates data insufficiency in CD. Notably, we achieve competitive results using only 20% training data than baseline (random) using 100% data. Both quantitative and qualitative results demonstrate the generalization ability of our pre-trained model to downstream images even remaining domain gaps with the pre-training data. Our Code will make public.
Abstract (translated)
URL
https://arxiv.org/abs/2205.13769
