Reproduce “The Change You Want to See” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ragav Sachdeva, Andrew Zisserman
Researchers from Oxford's Visual Geometry Group developed a framework that performs object-level change detection, effectively distinguishing true changes from geometric and photometric variations across image pairs. The co-attention-based Siamese U-Net, trained with scalable synthetic data, demonstrates robust zero-shot generalization across diverse real and synthetic datasets, including 3D camera motion.
Key insights
- Synthetically generated training data, created by inpainting and object pasting from existing datasets, can effectively train a robust change detection model, overcoming data scarcity.
- A co-attention mechanism is critical for implicitly establishing correspondences between images, allowing the model to handle significant geometric variations without explicit registration.
- Formulating change detection at the object-level with bounding box predictions simplifies the problem and enhances applicability for macroscopic changes, while being amenable to scalable data generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.