Incremental Object Detection Training with Retained Original Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models face challenges in adding new object classes without forgetting old classes, leading to catastrophic forgetting, and conventional methods like full retraining are time-consuming and resource-intensive.
Innovation Solution
A method involving data augmentation and automatic labeling using a combination of object detection and classifier models, along with low-rank adaptation, to add new object classes while retaining original weights and using random weights for new connections, thus maintaining performance on both old and new classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If full retraining is performed to add new object classes, then the model can detect new classes, but training time and computational resources increase significantly
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) training on old object classes to establish baseline performance, and (2) training on new object classes while preserving old class performance. This segmentation allows selective application of techniques like data augmentation and catastrophic forgetting prevention only during the second phase, reducing overall training time compared to full retraining.
Solution Approach 2:
The patent performs preliminary actions by pre-training the model on old object classes before introducing new classes. This preliminary training establishes a foundation that can be built upon, allowing the model to adapt to new classes more efficiently without starting from scratch, thus reducing the time required for adding new object classes.
2Adaptability or versatility
If full retraining is performed to add new object classes, then the model can detect new classes, but computational resources and costs increase
Solution Approach 1:
The patent applies partial action by using data augmentation techniques that generate synthetic training examples through transformations like rotation, flipping, and cropping. This creates additional training data without requiring proportional increases in computational resources for data collection and annotation, enabling efficient model adaptation to new object classes.
Solution Approach 2:
The patent changes training parameters such as learning rate, batch size, and training epochs when transitioning from old to new object classes. These parameter adjustments optimize the training process for incremental learning, reducing computational resource requirements compared to full retraining while maintaining effective model adaptation.
3Productivity
If training is performed on new object classes only, then training is efficient, but the model forgets old object classes (catastrophic forgetting)
Solution Approach 1:
The patent ensures continuity of useful action by implementing mechanisms that maintain performance on old object classes while training on new classes. This includes using data augmentation to generate diverse training examples that reinforce old class knowledge and employing training strategies that prevent catastrophic forgetting, thus maintaining continuous detection capability across all object classes.
Solution Approach 2:
The patent incorporates feedback mechanisms where the model's performance on old object classes is continuously monitored during training on new classes. This feedback allows for dynamic adjustment of training parameters and data sampling strategies to prevent performance degradation on old classes, ensuring reliable detection across the full range of object classes.
4Productivity
If data augmentation is applied to combine first and second images, then training efficiency improves, but image processing complexity increases
Solution Approach 1:
The patent uses copying by creating synthetic versions of existing images through data augmentation transformations. Instead of requiring new real-world images for each object class, the system generates augmented copies of existing images with various transformations, reducing the need for complex image acquisition and processing while improving training efficiency.
Data Source
Figure 1
Figure 2A
Figure 2B~2C
AI summary
Systems and methods are provided for implementing training of a pre-trained object detection model for detecting new object classes. In examples, to train an object detection model, which has been pre-trained with a first set of object classes, with a new object class, a computing system applies to each of a plurality of first images that each depicts an object corresponding to an object class among the first set of object classes, a set of data augmentations combining each first image with at least one second image among a plurality of second images that each depicts a second object corresponding to the new object class, to generate a plurality of augmented images. The computing system trains the object detection model using the plurality of augmented images. In examples, original weights corresponding to the first set of object classes are retained, while random weights are used for the new object class.