OOD Object Detection via Semantic Similarity Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) used in autonomous driving systems struggle to reliably detect out-of-distribution (OOD) objects, leading to inaccurate outputs and potentially unsafe driving behavior.
Innovation Solution
The proposed approach uses a class semantics-based anomaly criterion for OOD object detection, which involves determining the semantic similarity of object attributes across multiple sensor data sets using a semantic-aware distribution comparison metric.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DNN-based object detection is used for autonomous driving, then detection performance for known objects is improved, but reliability for out-of-distribution objects deteriorates
Solution Approach 1:
The patent segments the detection task into two independent parts: (1) standard object detection for known classes using DNN, and (2) OOD detection using temporal logit analysis. This segmentation allows each component to optimize for its specific function without interference, resolving the contradiction between detection accuracy for known objects and reliability for OOD objects.
Solution Approach 2:
The patent introduces class logits as an intermediary signal that bridges object detection and OOD detection. By analyzing temporal changes in class logits rather than directly detecting OOD objects, the system indirectly identifies anomalies while maintaining high performance for known object detection.
2Reliability
If standard OOD detection methods are used, then OOD detection capability is improved, but false positive rate increases due to uncertainties
Solution Approach 1:
The patent performs preliminary analysis of class logits during the normal object detection process before making OOD determination. By pre-computing and storing class logits for each detected object across multiple frames, the system prepares the necessary data for OOD detection without requiring additional computational resources or increasing false positives.
Solution Approach 2:
The patent implements feedback by comparing class logits across temporal frames and using this feedback to identify OOD objects. The system continuously monitors logit stability and uses deviations from expected patterns as feedback signals to flag potential OOD detections, thereby reducing false positives through iterative refinement.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the present disclosure relate to a vehicle, an apparatus, a computer program, and a method for detecting an out-of-distribution (OOD) object. The method comprises obtaining first sensor data of an object at a first point in time and second sensor data of the object at a second point in time. Further, the method comprises determining an attribute of the object using the first sensor data and determining the attribute of the object using the second sensor data. As well, the method provides for determining if the object is an OOD object based on a semantic similarity of the attribute for the first and the second sensor data.