Neural Network Training With Absolute Luminance for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional image processing pipelines in machine vision tasks, such as those used in self-driving vehicles and robotics, lack accurate measurement of luminance, leading to inaccurate training and inference in deep learning models, reducing their accuracy and reliability.
Innovation Solution
Incorporating calibration and calculation steps in the image signal processing pipeline to generate absolute luminance and radiance values for each pixel, enabling better object discrimination and improved light measurement across varying illumination conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If traditional image processing pipelines are used to yield aesthetically pleasing results, then visual quality is improved, but luminance measurement accuracy deteriorates
Solution Approach 1:
The image processing pipeline is segmented into multiple processing streams: one for aesthetic image enhancement and another for accurate luminance measurement. This allows each stream to be optimized for its specific purpose without compromising the other, resolving the contradiction between visual quality and measurement accuracy.
Solution Approach 2:
An intermediary luminance calculation module is introduced that operates on the original sensor data before aesthetic processing alters the luminance values. This intermediary preserves the raw luminance information needed for accurate measurement while allowing aesthetic processing to proceed separately for visual quality.
2Illumination intensity
If image processing focuses on aesthetic quality, then visual appeal is improved, but object detection accuracy deteriorates
Solution Approach 1:
The processing system is divided into separate aesthetic processing and object detection processing paths. The aesthetic path enhances visual appeal while the object detection path uses preserved raw luminance data to maintain detection accuracy, eliminating the trade-off between these two goals.
Solution Approach 2:
A copy of the original image data with accurate luminance values is created and preserved for object detection purposes, while the aesthetic processing is applied to a separate copy. This allows both aesthetic quality and detection accuracy to be maintained independently.
3Productivity
If training data lacks accurate luminance information, then processing speed is improved, but model robustness deteriorates
Solution Approach 1:
Accurate luminance information is pre-calculated and embedded in the training data during the data preparation phase. This preliminary action ensures that the neural network receives high-quality luminance data during training, improving model robustness without affecting inference speed since the luminance data is already prepared.
Solution Approach 2:
The system automatically generates and attaches accurate luminance values to training images through an integrated processing pipeline, eliminating the need for manual annotation of luminance data. This self-service approach maintains processing efficiency while ensuring high-quality training data with accurate luminance information.
Data Source
AI summary
Apparatuses, systems, and techniques to process luminance and/or radiance values of one or more images from one or more cameras using one or more neural networks to perform a machine vision task. In at least one embodiment, one or more neural networks determine detection difficulty levels of objects within the one or more images and performs a machine vision task based on the determined detection difficulty levels of objects within images associated with that ask.