3DCD: Scene Independent End-to-End Spatiotemporal Feature Learning Framework for Change Detection in Unseen Videos.

Area of Science:

• Computer Vision
• Video Processing
• Deep Learning

Background:

• Supervised deep learning methods dominate change detection but require annotated training data.
• Their performance on unseen videos (scene-independent) is largely undocumented.
• Existing methods lack generalized models for real-world change detection applications.

Purpose of the Study:

• To introduce a scene-independent evaluation (SIE) framework for generalized change detection models.
• To propose a fast, lightweight, and end-to-end 3D-CNN based change detection network (3DCD).
• To compare the proposed 3DCD network against state-of-the-art methods in both SIE and scene-dependent evaluation (SDE) setups.

Main Methods:

• Developed a scene-independent evaluation (SIE) framework for assessing model generalization.
• Proposed a lightweight, end-to-end 3D-CNN model (3DCD) with spatiotemporal learning blocks.
• Incorporated background estimation, motion saliency, and multi-schematic feature encoding-decoding for foreground segmentation.

Main Results:

• The proposed 3DCD network achieved state-of-the-art performance in both SIE and SDE evaluations.
• Demonstrated superior results on benchmark datasets: CDnet 2014, LASIESTA, and SBMI2015.
• The 3DCD model is fast (25 fps) and lightweight (1.16 MB).

Conclusions:

• The 3DCD network provides a generalized solution for change detection, effective even on completely unseen videos.
• The study establishes clear SIE and SDE evaluation protocols for change detection research.
• The proposed framework and model advance the field by addressing the limitations of existing supervised methods.