Latent bridge matching improves albedo estimation with shading conditioning
Albedo Estimation via Latent Bridge Matching
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Albedo estimation is about figuring out the true color of objects in images, separate from lighting effects. The authors propose a new method called latent bridge matching (LBM) to improve this estimation by making the results more physically consistent, faster to compute, and better at handling different types of images. They enhance their method by conditioning on shading information and making shading prediction depend on estimated albedo, which helps create more accurate image reconstructions. Their approach performs well compared to other state-of-the-art methods on several datasets.
What this means in practice
- •For computer vision engineers: Create more accurate and physically consistent models for separating lighting and object colors in images.
- •For augmented reality developers: Produce better albedo maps for virtual object integration that adapt across different lighting conditions in real-time.
Authors
Carme Corbi, David Serrano-Lozano, Javier Vazquez-Corral, Maria Vanrell
Abstract
Recent advances in Intrinsic Image Decomposition (IID) have increasingly relied on generative models. However, progress remains limited by three key challenges: (a) insufficient physical consistency, (b) high computational cost at inference time, and (c) limited generalization capabilities. In this work, we show that latent bridge matching (LBM) effectively addresses these limitations for albedo estimation. We introduce a novel LBM-based architecture that enforces physical consistency through a pixel reconstruction loss, benefits from the inherent efficiency of LBM low-cost inference, and improves generalization across diverse datasets by incorporating a shading conditioning. In this extended version, we additionally show that conditioning the shading estimator itself on the predicted albedo further improves reconstruction fidelity, and we benchmark our best model against stateof-the-art IID methods across five real and synthetic datasets.