Radiative Backpropagation with Non-Static Geometry

2025-04-08

AI Summary

Radiative backpropagation (RB) assumes static geometry, leading to biased gradients under dynamic geometry due to detached sampling—an artifact of implicitly enforcing this static assumption. Method: The authors derive, for the first time, a general radiative gradient formula for non-static geometry, explicitly characterizing how geometric motion affects the derivative of the path integral. Building on this theory, they propose two equivalent unbiased backpropagation formulations: attached path replay and surface reparameterization—both rigorously satisfying the differentiability–integrability interchange condition. Contribution/Results: These methods eliminate gradient bias induced by geometric motion. Experiments on simple dynamic scenes empirically validate the bias in conventional RB and demonstrate that the proposed approaches recover unbiased gradients. The work establishes a more rigorous theoretical foundation for differentiable rendering and provides practical, provably correct implementations for dynamic scenes.

Technology Category

• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Intelligent Robots: Motion and Path Planning
• Computer Vision: Motion & Tracking

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

One of the core working principles of Radiative Backpropagation (RB) is that differential radiance is transported like normal radiance. This report shows that this is only true if scene geometry is static. We suggest that static geometry is an implicit assumption in the current theory, leading to biased gradients in implementations based on detached sampling, and demonstrate this with simple examples. We derive the general derivatives for non-static geometry: the RB-based derivatives with detached sampling are obtained either by an algorithm similar to attached path replay backpropagation or by a construction that reparameterizes the rendering integral over surfaces.

Problem

• Identifies bias in Radiative Backpropagation with dynamic geometry
• Proposes solutions for non-static geometry derivatives
• Addresses gradient inaccuracies in detached sampling implementations

Innovation

• Non-static geometry derivatives for RB
• Attached path replay backpropagation algorithm
• Reparameterization of rendering integral surfaces