ROI-Guided Unknown Object Labeling for Automated Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition methods in highly automated driving fail to reliably identify unknown objects, leading to potential hazards due to low confidence levels, which can result in overreactions or underreactions by the vehicle.
Innovation Solution
A method involving a pretrained segmentation model that preprocesses the image by defining a Region of Interest (ROI) to focus on unknown objects, using a mask segmentation model to generate an Out-of-Distribution (OOD) mask and morphing the ROI to eliminate gaps, followed by false-positive removal, ensuring accurate labeling of unknown objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the pretrained segmentation model processes the whole image, then all objects can be labeled, but computing power is wasted and information about unknown objects is lost
Solution Approach 1:
The image processing is segmented into two stages: first, a mask segmentation model identifies regions containing unknown objects; second, the pretrained segmentation model processes only those specific regions. This segmentation approach prevents the pretrained model from wasting computational resources on already-known objects while ensuring unknown objects are properly labeled.
Solution Approach 2:
The mask segmentation model performs preliminary action by pre-identifying regions containing unknown objects before the pretrained segmentation model processes the image. This preliminary filtering step ensures that only relevant regions are passed to the computationally intensive pretrained model, optimizing both accuracy and resource usage.
2Loss of information
If the pretrained segmentation model processes the whole image, then all objects can be labeled, but information about which objects were initially unknown is lost
Solution Approach 1:
The processing pipeline is segmented into distinct stages: mask generation, region filtering, and targeted segmentation. Each stage preserves and passes forward the information about which objects are unknown, ensuring this critical information is not lost while maintaining manageable processing complexity through modular design.
Solution Approach 2:
The system implements feedback by using the mask segmentation model's output to guide the pretrained segmentation model's processing. The information about unknown objects detected in the first stage feeds back into the second stage, ensuring that the pretrained model focuses its computational power on regions that need labeling while preserving information about object status.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A computer-implemented method (100) for labeling unknown objects (150) in an image (104) comprises a step of receiving (102) the image (104), a step of determining (148) a first Region of Interest (134), whereas the first Region of Interest (134) is a sample of the image (104) comprised of the unknown objects (150), and a step of labeling the unknown objects (150) using a pretrained segmentation model (146) based on the first Region of Interest (134).