Hybrid Multi-Box Loss for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited availability of completely annotated training data sets hampers the effective training of neural network-based object classifiers, particularly for multi-class object classifiers, as existing methods like negative mining can lead to inaccurate results when expanding the classifier's functionality with partially annotated data.
Innovation Solution
A method and system that employs a hybrid multi-box loss function to train neural networks using a mixed dataset, categorizing data into types where all classes are marked, some classes are marked, and none are marked, allowing for the use of partially annotated data without compromising accuracy by adjusting loss calculations and incorporating Type 3 data to provide sufficient negative examples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If negative mining is used to train a classifier with partially annotated data, then training speed and efficiency are improved, but classification accuracy deteriorates due to inaccurate background assumptions
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different data types within the training set. Type 1 images (fully annotated) undergo negative mining with all background regions treated as negative samples, while Type 2 images (partially annotated) apply negative mining only to regions outside the annotated bounding boxes. This localized application of negative mining preserves training efficiency while maintaining accuracy by avoiding false negative assumptions in partially annotated data.
2Measurement precision
If fully annotated training data is used, then classification accuracy is improved, but data preparation time and cost increase significantly
Solution Approach 1:
The patent applies partial action by accepting that not all regions in partially annotated images need to be correctly classified. For Type 2 images, the loss function is modified to only enforce positive classification for annotated objects, while unannotated regions are excluded from negative mining. This partial application of annotation requirements reduces data preparation burden while maintaining sufficient training effectiveness.
3Adaptability or versatility
If the number of object classes to be classified increases, then classifier functionality is improved, but the difficulty and cost of acquiring completely annotated training data increases
Solution Approach 1:
The patent creates a universal training framework that handles multiple data types (Type 1 fully annotated, Type 2 partially annotated, Type 3 unannotated) and multiple object classes simultaneously. The hybrid loss function and conditional negative mining approach work consistently across different annotation completeness levels and any number of object classes, enabling the classifier to be extended to new classes without requiring complete re-annotation of the entire dataset.
Data Source
AI summary
A method for training a neural network for detecting a plurality of classes of object within a sample comprises providing a training data set comprising a plurality of samples, each annotated according to whether the samples include labelled objects of interest. In a first type of samples, all objects of interest are labelled according to their class and comprise a foreground of the samples, the remainder of the samples comprising background. In a second type of samples, some objects of interest are labelled in a foreground and their background may comprise unlabelled objects. A third type of samples comprise only background comprising no objects of interest. Negative mining is only performed on the results of processing the first and third types of samples.