Cross-Sensor Object Recognizer Training With Error-Guided Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object recognition technologies using sensors like cameras, lidar, or radar require significant training data and time, which is costly and inefficient.
Innovation Solution
A method and device for training an object recognizer that utilizes a first sensor-based object recognizer and a second sensor-based object recognizer, with an error detection mechanism to identify errors in the second sensor-based recognizer, and a regression analyzer to estimate a predicted value for training, thereby reducing the need for extensive training data and time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a new sensor-based object recognizer is trained from scratch using conventional methods, then the object recognition accuracy can be improved, but the training time and costs increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training an object recognizer using data from a first sensor type before fine-tuning it with data from a second sensor type. This pre-training phase establishes foundational recognition capabilities, allowing the subsequent sensor-specific training to focus on adapting to the new sensor's characteristics rather than learning from scratch, thereby significantly reducing training time while maintaining accuracy
Solution Approach 2:
The patent uses copying by transferring the knowledge and weights from a pre-trained object recognizer (trained on first sensor data) to create an initial model for the second sensor-based recognizer. This copied model serves as a starting point that already contains general object recognition patterns, eliminating the need to train all parameters from random initialization and thus reducing both time and computational costs
2Measurement precision
If a new sensor-based object recognizer is trained from scratch using conventional methods, then the object recognition accuracy can be improved, but the training costs increase significantly
Solution Approach 1:
The pre-training phase on first sensor data performs the expensive computation of learning general object recognition patterns once, and this pretrained model is then reused for the second sensor. This eliminates the need to perform the same expensive computation again from scratch, significantly reducing training costs while maintaining the ability to achieve high accuracy on the new sensor
Solution Approach 2:
The patent changes parameters by adjusting the training approach from full training from random initialization to fine-tuning from a pretrained state. This parameter change in the training strategy reduces computational resources required, lowering costs while still achieving the necessary accuracy through targeted fine-tuning on the new sensor data
3Adaptability or versatility
If conventional object recognition training is used for new sensors, then comprehensive object recognition capability can be achieved, but the training data requirements become excessive
Solution Approach 1:
The pre-training on first sensor data provides a comprehensive foundation for object recognition that transfers to the second sensor. This preliminary learning of general patterns reduces the amount of second sensor data needed, as the model already understands object categories and features from the pre-training phase and only needs to adapt to the new sensor's specific characteristics
Data Source
AI summary
The disclosure relates to an object recognizer training method and device. There may be provided a method for training an object recognizer comprising obtaining an image by capturing an object by a first sensor and a second sensor, obtaining first object recognition information by inputting an image captured by the first sensor to a first sensor-based object recognizer and obtaining second object recognition information by inputting an image captured by the second sensor to a second sensor-based object recognizer, detecting an object recognition error in the second sensor-based object recognizer, if the object recognition error is detected, obtaining a predicted value of the second object recognition information corresponding to the first object recognition information based on reference data created before, and training the second sensor-based object recognizer using the predicted value of the second object recognition information.