Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/ragavsachdeva/CYWS-3D/llms.txt

Use this file to discover all available pages before exploring further.

CYWS-3D is a deep learning system for change detection between two images of the same scene. Given a pair of RGB images—optionally with depth maps and camera parameters—it predicts bounding boxes around regions that have changed, in both images simultaneously. The model was presented at ICCVW 2023 by Ragav Sachdeva and Andrew Zisserman.

Installation

Set up the Conda environment, install all dependencies, and download the pre-trained checkpoint.

Quickstart

Run batched inference on example image pairs in minutes using the pre-trained model.

Registration Strategies

Understand the 3D, 2D affine, and identity registration modes and when to use each.

Model Reference

Explore the Model class, its constructor arguments, and public methods.

​
How it works

CYWS-3D aligns two views of a scene in feature space before comparing them, so viewpoint and lighting changes do not confuse the detector. The alignment strategy adapts to the information available:
1

Provide an image pair

Supply two RGB images (image1, image2) and choose a registration_strategy: 3d, 2d, or identity.
2

Extract depth and correspondences

For 3D scenes the pipeline runs ZoeDepth (if no depth is provided) and SuperPoint + SuperGlue to find correspondences between the images.
3

Warp features into alignment

The Feature Registration Module projects image features into a common 3D space (or applies a 2D affine warp) and renders the warped features back to image resolution.
4

Detect changed regions

A DINO ViT-Base backbone, UNet decoder, and CenterNet head predict bounding boxes around changed regions in each image. Post-processing filters low-area boxes, suppresses overlapping boxes, and matches pairs across images.

​
Key capabilities

No poses required

Run with RGB images only — the model estimates depth and correspondences automatically.

Mixed batches

A single batch can contain samples with different registration strategies (3D, 2D, identity) processed together.

Two benchmark datasets

KC-3D and RC-3D datasets with ground-truth depth, camera parameters, and change annotations are available for download.

Flexible post-processing

Control confidence thresholds, minimum bbox area, NMS overlap, and cross-image matching from a single inference call.

​
Citation

If you use CYWS-3D in your research, please cite: 
@InProceedings{Sachdeva_ICCVW_2023,
    title = {The Change You Want to See (Now in 3D)},
    author = {Sachdeva, Ragav and Zisserman, Andrew},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year = {2023},
}

Build docs developers (and LLMs) love

Get started for freeTalk to us