Vehicle Image Annotation Using Motion-Tracked Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep neural networks (DNNs) for object detection in images requires large datasets with accurate annotations, but unreliable annotations can lead to poor performance, especially in scenarios like vehicle guidance where anomalous data detection is critical.
Innovation Solution
A method that uses a DNN to annotate images acquired by a vehicle's sensors, translates bounding boxes based on latitudinal and longitudinal motion data, and applies image processing techniques like super resolution, blurring, and hierarchical pyramid processing to improve annotation accuracy and generate a larger, more reliable training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a DNN is trained with a limited training dataset, then training time and computational resources are reduced, but the DNN fails to detect objects in new environments or under different conditions
Solution Approach 1:
The system performs preliminary actions by using a first DNN to generate bounding boxes for objects in images before the actual training process. These pre-generated annotations serve as initial training data, enabling the second DNN to be trained more efficiently while maintaining good detection performance across different environments.
Solution Approach 2:
The patent introduces an intermediary process where a first DNN acts as a mediator to generate bounding box annotations. This intermediary DNN bridges the gap between having no annotations and requiring extensive manual annotations, allowing the system to train a second DNN with sufficient training data without manual labeling efforts.
2Measurement precision
If manual annotation of images is performed to create training datasets, then annotation accuracy is improved, but time consumption and labor costs increase significantly
Solution Approach 1:
The system implements self-service by using the first DNN to automatically generate bounding box annotations for the training dataset. Instead of relying on manual annotation, the DNN annotates images itself, eliminating the need for human annotators while maintaining consistent and reliable annotation quality across the entire dataset.
Solution Approach 2:
The patent uses copying by having the first DNN replicate the annotation process across multiple images. The bounding box generation capability is copied and applied systematically to create a comprehensive training dataset, replacing manual copying and pasting of annotations with automated DNN-based replication.
3Measurement precision
If the DNN is trained with diverse training data including various poses and sizes, then detection accuracy across different scenarios is improved, but the complexity of data collection and annotation increases
Solution Approach 1:
The first DNN performs multiple functions: it detects objects, determines their locations, estimates sizes, and generates bounding boxes. This multi-functional approach eliminates the need for separate processes for each annotation task, reducing overall system complexity while enabling comprehensive training data generation with various poses and sizes.
Solution Approach 2:
The system leverages parameter changes by utilizing the first DNN's ability to detect and annotate objects with varying parameters such as position, size, and orientation. By automatically adapting to different object parameters in the images, the system generates diverse training data without increasing annotation complexity, as the DNN dynamically adjusts its annotations based on detected object characteristics.
Data Source
AI summary
A first image can be acquired from a first sensor included in a vehicle and input to a deep neural network to determine a first bounding box for a first object. A second image can be acquired from the first sensor. Input latitudinal and longitudinal motion data from second sensors included in the vehicle corresponding to the time between inputting the first image and inputting the second image. A second bounding box can be determined by translating the first bounding box based on the latitudinal and longitudinal motion data. The second image can be cropped based on the second bounding box. The cropped second image can be input to the deep neural network to detect a second object. The first image, the first bounding box, the second image, and the second bounding box can be output.