Scene Parameter Saliency via Differentiable Light Transport

2026-07-23

AI Summary

This work proposes a differentiable rendering–based method for scene parameter saliency analysis, aiming to quantify the influence of scene parameters on specific scalar metrics in rendered images—such as glare index, mean luminance, or neural perceptual scores. By integrating reverse-mode automatic differentiation with multi-bounce light transport simulation, the approach pioneers the use of differentiable rendering to generate “metric saliency maps,” enabling direct attribution within the physical light transport process and overcoming limitations of conventional neural-network-based saliency methods. Experiments demonstrate that the resulting saliency distributions are strongly dependent on the target metric itself rather than solely on the intrinsic scene structure, thereby validating the informational value and efficacy of derivative images for interpretable scene understanding.

Abstract

Gradient-based saliency methods reveal which input features most influence a neural network's output, and are a standard tool for model interpretability. We observe that differentiable renderers, which are conventionally used for parameter optimisation, produce an analogous form of saliency: given any scalar metric evaluated on a rendered image, a single reverse-mode differentiation pass yields per-parameter gradients that identify which scene elements most influence the metric. We call these gradient fields metric saliency maps. Unlike neural saliency, which propagates attribution through learned weights, metric saliency propagates through the image formation process itself, including multi-bounce light transport, capturing parameter dependencies that are semi-opaque to manual inspection. We compute metric saliency maps for qualitatively different objectives: psychovisual glare indices, mean scene luminance, and neural perceptual scores. The saliency rankings differ substantially across metrics for the same scene, with parameters that dominate one objective being negligible for another. The saliency map is specific to the metric, not an intrinsic property of the scene. Our results suggest that differentiable renderers produce derivative images that are as informative for scene understanding as the primal images they were designed to generate.

Problem

• scene parameter saliency
• differentiable rendering
• light transport
• metric sensitivity
• image formation

Innovation

• differentiable rendering
• metric saliency
• light transport
• gradient-based interpretability
• scene understanding