Scene Change Detection

Abstract

Scene Change Detection (SCD) identifies changes between the images taken at two different times, using pixel-level or point cloud approaches in most cases. Training a neural network for such a task requires a large number of images with annotated changes. Annotating changes is a slow, costly and time-consuming process. The state-of-the-art (SOTA) approach for SCD, like the DR-TANet paper, is based on transfer learning from large ImageNet datasets. This is a supervised technique and to overcome the challenges mentioned above, we introduce a self-supervised pretraining method with unlabeled datasets based on a existing D-SSCD approach that learns temporal-consistent representations of a pair of images. This project is an investigation of these approaches that can train and evaluate on available datasets through the use of a suitable loss function for the purpose of SCD. We compare results for different percentages of labeled data from different models and benchmark datasets such as Visual Localization CMU (VL_CMU_CD) and Panoramic change detection (PCD) datasets.