Ray Sample Reconstruction for Immediate DNN Relearning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing rendering systems using deep neural networks (DNNs) for ray-tracing face challenges in maintaining performance when data characteristics differ between training and operation, leading to increased calculation costs and residual noise due to mismatched learning coefficients, especially in applications requiring immediate updates and limited computational resources.
Innovation Solution
A method to generate learning data without additional rendering by reconstructing ray tracing data to create teacher and student images with adjustable resolution and spp values, using accumulated ray sample data from viewport videos, thereby reducing calculation costs and enabling immediate learning updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a teacher image of high sample per pixel (spp) is used for relearning, then the learning performance is improved, but the calculation cost for rendering increases
Solution Approach 1:
The patent performs preliminary action by accumulating ray sample data during the normal rendering process before learning is needed. This accumulated data is stored and can be reused for generating teacher images at any time without requiring additional rendering calculations, thus resolving the contradiction between learning performance and calculation cost
Solution Approach 2:
The patent creates copies of ray sample data at different accumulation levels to generate both low-spp student images and high-spp teacher images from the same original rendering data. This copying approach allows the system to obtain high-quality teacher images without performing expensive high-spp rendering, as the data is already available from the initial rendering pass
2Manufacturing precision
If ray sample data is accumulated to generate high-spp teacher images, then image quality is improved, but rendering time increases
Solution Approach 1:
The system performs preliminary accumulation of ray sample data during the standard rendering process. This data is stored in memory or storage and can be accessed later for generating teacher images of any desired quality level without requiring additional rendering time, thus decoupling image quality from rendering time
Solution Approach 2:
The patent transitions from the traditional dimension where image quality is directly tied to rendering parameters (spp, resolution) to a new dimension where quality is determined by data accumulation level. By storing intermediate ray sample data, the system can generate images at any quality level from the same rendering pass, adding a temporal dimension to quality control
Data Source
AI summary
An information processing device includes a learning data acquisition processing unit. The learning data acquisition processing unit sequentially acquires, from a ray tracer, ray sample data generated by ray simulation by the ray tracer. The learning data acquisition processing unit reconstructs the ray sample data sequentially acquired from the ray tracer, and generates learning data of an inference model. The learning data includes a student image and a teacher image for learning the inference model.