Quantile Neural Network for Object Detection Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) used in image sensor-based machine learning systems for object detection in vehicles and robots face challenges with inaccurate predictions, especially for new data and outlier data due to lack of knowledge and noise factors, leading to degraded system performance.
Innovation Solution
A quantile neural network is trained to output confidence or uncertainty values corresponding to prediction outputs, allowing decision controllers to assess the reliability of predictions and determine whether to reject or use them, thereby improving performance by estimating both aleatoric and epistemic uncertainties using a single model and inference, reducing computational resource consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep neural networks are used for object detection in image data, then object detection capability is improved, but computational resource consumption increases
Solution Approach 1:
The patent segments the object detection task into multiple quantile levels (e.g., 0.1, 0.3, 0.5, 0.7, 0.9 quantiles), where each quantile network is trained to detect objects at different confidence levels. This segmentation allows the system to process images at multiple resolution levels, improving detection accuracy while reducing overall computational load by avoiding full-resolution processing for all detection tasks.
Solution Approach 2:
The patent implements dynamic adjustment of quantile levels based on image characteristics and detection requirements. The system can adaptively select which quantile networks to apply to different images or regions, dynamically optimizing the balance between detection accuracy and computational resource consumption rather than uniformly applying all quantile levels to every image.
2Reliability
If multiple quantile levels are processed to improve prediction reliability, then confidence assessment is improved, but device complexity increases
Solution Approach 1:
The patent merges multiple quantile detection results into a unified confidence assessment mechanism. Instead of treating each quantile level as a separate complex system, the results from different quantile networks are combined to produce overall prediction confidence scores, reducing the effective complexity of the multi-quantile system while maintaining reliable confidence assessment.
Solution Approach 2:
The quantile neural network structure is designed to be universal across different quantile levels, where a single base network architecture is reused and adapted to multiple quantile requirements. This multi-functionality reduces device complexity by avoiding the need for entirely separate network structures for each quantile level, instead using one versatile framework that handles all quantile assessments.
Data Source
AI summary
A computer, including a processor and a memory, the memory including instructions to be executed by the processor to train a quantile neural network to input an image and output a lower quantile (LQ) prediction, a median quantile (MQ) prediction and an upper quantile (UQ) prediction corresponding to an object in the image, wherein an LQ loss, an MQ loss and a UQ loss are determined for the LQ prediction, the MQ prediction and the UQ prediction respectively and wherein the LQ loss, the MQ loss and the UQ loss are combined to form a base layer loss and output the quantile neural network.