CYWS-3D (The Change You Want to See, Now in 3D) is a PyTorch deep learning model that detects meaningful semantic changes between two RGB images of the same 3D scene taken from different viewpoints. Unlike classical change-detection approaches that assume a fixed camera, CYWS-3D explicitly accounts for the 3D geometry of the scene: it estimates or accepts depth maps, registers the two views in 3D space, and then identifies objects or regions that have genuinely appeared, disappeared, or been altered — not merely moved in the image plane due to parallax. The model was presented at ICCVW 2023 and builds directly on the 2D CYWS framework, extending it to handle real-world, multi-viewpoint capture scenarios.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.
Installation
Set up the conda environment, install all dependencies, and download the pre-trained checkpoint.
Quickstart
Run inference on the bundled demo images in minutes with a single command.
Concepts: Overview
Understand the architecture, the 3D registration pipeline, and how change detection works end-to-end.
Registration Strategies
Learn when to use the
3d, 2d, or identity registration strategy for your data.What it detects
CYWS-3D is designed for the semantic change-detection task: given two RGB images of the same physical location captured at different times (or from different viewpoints), the model produces bounding-box predictions around regions where the scene content has genuinely changed. A typical use case is comparing a “before” photograph to an “after” photograph of a room, building interior, or outdoor area — the model highlights objects that have been added, removed, or significantly modified. Because the two images are taken from different positions in 3D space, naive pixel-level comparison fails — parts of the background shift in the frame purely due to parallax, producing false positives. CYWS-3D solves this by warping the images into alignment using estimated or provided depth and camera geometry before applying the change-detection head, so only true scene-level changes survive.Key capabilities
No Camera Poses Required
The model runs on a plain pair of RGB images with no metadata. Depth is predicted automatically using ZoeDepth, and relative camera pose is estimated internally.
Three Registration Strategies
Choose
3d for real perspective-shifted scenes, 2d for planar/affine-transformed image pairs, or identity when the images are already aligned and no warping is needed.Automatic Depth Prediction
When depth maps are not supplied, ZoeDepth is invoked automatically to produce metric depth estimates, enabling full 3D registration without any extra tooling.
Batched Mixed-Strategy Inference
A single
input_metadata.yml file can contain items that use different registration strategies in the same batch, letting you process heterogeneous datasets in one forward pass.