3D Object Detection Using Pseudo-Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and time-intensive process of obtaining large labeled datasets for training autonomous vehicle perception systems, especially for new deployment geographies, limits the efficacy of onboard perception systems in diverse environments.
Innovation Solution
The use of pseudo-labeling for 3D object detection, which leverages unlabeled data to train computationally efficient student neural networks, allowing for effective performance in new domains with reduced labeled dataset requirements and improved latency and compute efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large labeled datasets are collected to improve perception system accuracy, then detection accuracy is improved, but data acquisition cost and time increase significantly
Solution Approach 1:
The patent uses pseudo-labels as synthetic copies of ground truth annotations. A teacher model generates pseudo-labels from unlabeled data, creating artificial training labels that replicate the function of expensive manual annotations without requiring human time investment, thus resolving the contradiction between accuracy and data acquisition time
Solution Approach 2:
The patent performs preliminary labeling of unlabeled data using a teacher model before training the student model. By pre-generating pseudo-labels from available unlabeled data, the system prepares training materials in advance without requiring time-intensive manual annotation during the actual training phase
2Measurement precision
If computationally expensive teacher models are used to improve detection accuracy, then detection accuracy is improved, but computational efficiency and latency worsen
Solution Approach 1:
The student model is trained to copy the detection capabilities of the teacher model using pseudo-labels. This creates a lightweight replica that achieves similar accuracy to the expensive teacher model but with much lower computational requirements and latency, resolving the contradiction between accuracy and computational efficiency
Solution Approach 2:
The patent replaces the expensive, complex teacher model with a cheaper, simpler student model for actual deployment. The student model uses fewer computational resources and can be executed efficiently on autonomous vehicle hardware, sacrificing some model complexity to gain computational efficiency
3Manufacturing precision
If manual 3D bounding box annotations are used to label data, then labeled dataset quality is improved, but annotation cost and time increase
Solution Approach 1:
The patent replaces manual human annotation with automated pseudo-label generation by the teacher model. This creates synthetic annotations that mimic the quality of manual labels without requiring human annotators, dramatically reducing both time and cost while maintaining acceptable annotation quality for training purposes
Solution Approach 2:
The system performs self-annotation by using the teacher model to automatically generate pseudo-labels from unlabeled data. This eliminates the need for external human annotation services, making the labeling process autonomous and significantly reducing dependency on expensive manual labor
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network to perform 3D object detection. One of the methods includes training a student neural network to perform 3D object detection using pseudo-labels generated by a teacher neural network.