DNN Confidence Valves via K-Means Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep neural networks (DNNs) in vehicle advanced driver-assistance systems (ADAS) and autonomous driving often produce untrustworthy high confidence values due to limited training datasets, leading to inaccurate object detection and sensor fusion outputs.
Innovation Solution
The proposed solution involves K-means clustering of training histograms to determine cluster centers and adjust confidence scores based on distance ratios using a scaled sigmoid function, which calculates discount probability values to refine the confidence scores for improved object detection and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DNNs are trained using limited training datasets, then the system can be kept simple and fast, but the DNNs produce untrustworthy high confidence values for scenarios not in training data
Solution Approach 1:
The patent performs preliminary action by clustering training data into K clusters and computing cluster centers before actual inference. This pre-computed clustering structure allows the system to quickly evaluate confidence values during runtime without retraining, resolving the contradiction between fast training and reliable confidence assessment for unseen scenarios.
Solution Approach 2:
The patent introduces an intermediary mechanism using K-means clustering as a mediator between the limited training data and the inference process. The cluster centers serve as intermediaries that enable the system to generalize confidently to unseen scenarios by comparing input data to these pre-computed cluster representations, thereby improving reliability without requiring extensive training data.
2Reliability
If DNNs are trained on comprehensive datasets, then confidence values become more reliable, but the training process becomes more complex and time-consuming
Solution Approach 1:
The patent applies segmentation by dividing the training data into K distinct clusters based on feature space characteristics. This segmentation allows the system to manage comprehensive training data in an organized manner, computing cluster centers that represent each segment. During inference, the system only needs to compare inputs to these K cluster centers rather than processing the entire comprehensive dataset, thereby maintaining reliability while reducing computational complexity.
3Device complexity
If conventional sensor fusion is used with DNN outputs, then the system can operate with simple processing, but inaccurate confidence values lead to incorrect sensor fusion results
Solution Approach 1:
The patent implements feedback by using the K-means clustering results to adjust and refine the confidence values output by the DNN before they are used in sensor fusion. The system compares input data to cluster centers and uses this feedback information to modulate confidence scores, ensuring that only accurate and trustworthy confidence values are passed to the sensor fusion module. This feedback mechanism maintains processing simplicity while significantly improving measurement precision and object detection accuracy.
Data Source
AI summary
Vehicle perception techniques include obtaining a training dataset represented by N training histograms, in an image feature space, corresponding to N training images, K-means clustering the N training histograms to determine K clusters with respective K respective cluster centers, wherein K and N are integers greater than or equal to one and K is less than or equal to N, comparing the N training histograms to their respective K cluster centers to determine maximum in-class distances for each of K clusters, applying a deep neural network (DNN) to input images of the set of inputs to output detected/classified objects with respective confidence scores, obtaining adjusted confidence scores by adjusting the confidence scores output by the DNN based on distance ratios of (i) minimal distances of input histograms representing the input images to the K cluster centers and (ii) the respective maximum in-class.