Synthetic CNN Training Images for Rare Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to collect training images for CNNs that include unique objects rarely found in real driving situations, such as humans or animals, which hinders effective detection in autonomous driving scenarios.
Innovation Solution
A method is developed to generate synthesized images by combining original road images with additional labels for specific objects using a first CNN module, and then refine these images through a second CNN module to enhance their realism and discernibility, incorporating bounding box information to improve training data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training images are collected from normal driving video data, then common objects (cars, bicycles, humans) can be learned, but unique objects rarely found on roads (tigers, alligators) cannot be easily acquired
Solution Approach 1:
The patent uses image synthesis technology to create artificial training images by copying and combining real object images with road background images. This allows unique objects like tigers and alligators to be introduced into training data without requiring actual collection from driving videos, thus resolving the contradiction between detection capability and data collection ease
Solution Approach 2:
The patent introduces an image synthesis system as an intermediary between real driving data and training requirements. This intermediary generates synthetic training images that bridge the gap between available real-world data and the need for diverse unique objects, enabling CNN training without direct collection of rare objects
2Adaptability or versatility
If training images include a variety of unique objects, then CNN can learn to detect rare objects, but the complexity of data collection and processing increases
Solution Approach 1:
Instead of manually collecting and processing diverse real-world images of unique objects, the system copies existing object images and synthetically combines them with road backgrounds. This automated copying and composition process reduces the complexity of data preparation while maintaining variety in training objects
Solution Approach 2:
The patent performs preliminary image synthesis to pre-generate training images with unique objects before actual CNN training begins. This preliminary action prepares diverse training data in advance, eliminating the need for complex real-time data collection and processing during model development
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of generating at least one image data set to be used for learning CNN capable of detecting at least one obstruction in one or more autonomous driving circumstances, comprising steps of: (a) a learning device acquiring (i) an original image representing a road driving circumstance and (ii) a synthesized label obtained by using an original label corresponding to the original image and an additional label corresponding to an arbitrary specific object, wherein the arbitrary specific object does not relate to the original image; and (b) the learning device supporting a first CNN module to generate a synthesized image using the original image and the synthesized label, wherein the synthesized image is created by combining (i) an image of the arbitrary specific object corresponding to the additional label and (ii) the original image.