Deep Learning Classification Training With Soft Overlap Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional binary labelling in deep learning for object detection and classification leads to inefficiencies and inaccuracies in determining internal adjustable parameters, resulting in reduced precision and accuracy during the training process.
Innovation Solution
Implementing a soft labelling technique that assigns probability values between 0 and 1 based on the overlap area between detection and ground-truth bounding boxes, allowing for a weighted sum of softmax probabilities and proper back-propagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If binary labelling is used for classification, then the training process is simple and fast, but the precision and accuracy of determining internal adjustable parameters deteriorates
Solution Approach 1:
The patent changes the parameter of label values from binary (0 or 1) to continuous probability values (0 to 1). This allows the classification labels to represent degrees of membership rather than absolute categories, enabling more precise gradient computation and parameter updates during back-propagation while maintaining training efficiency through the same network architecture and loss function framework.
Solution Approach 2:
The patent introduces dynamic label assignment where labels are not fixed binary values but are computed as probability values based on overlap areas between detection boxes and ground truth. This dynamic approach allows the training process to adaptively adjust label values according to the degree of match, improving parameter determination precision without requiring a complete redesign of the training pipeline.
2Measurement precision
If soft labelling with probability values is implemented, then the accuracy of determining internal adjustable parameters is improved, but the complexity of the training process increases
Solution Approach 1:
The system performs self-service by automatically computing soft labels through overlap area calculations between detection boxes and ground truth boxes. This eliminates the need for manual label assignment or complex external annotation tools, as the system generates its own training labels based on geometric relationships, thereby improving precision without proportionally increasing operational complexity.
Solution Approach 2:
The overlap area calculation serves as an intermediary mechanism that bridges the gap between simple binary classification and complex precise parameter determination. By introducing this intermediate geometric measurement, the system achieves accurate soft label generation through a relatively simple area-based computation rather than requiring complex probabilistic models or additional processing stages.
3Use of energy by moving object
If binary classification labels are used, then the training process is computationally efficient, but the detection accuracy and reduction of missed detections deteriorates
Solution Approach 1:
The patent changes the label parameter from binary to continuous probability values, allowing the model to learn nuanced boundaries between classes. This enables more accurate detection by representing partial overlaps and ambiguous cases with intermediate probability values rather than forcing binary decisions, thereby improving detection accuracy while maintaining computational efficiency through the same forward-propagation and back-propagation framework.
Data Source
AI summary
Methods and apparatuses for deep learning training are provided which include receiving a candidate unit for classification. The candidate unit including an intersection area between a ground-truth bounding box and a detection box. The candidate unit is classified by assigning a label that is a probability value that a given feature is observed in the intersection area. Deep learning training is performed using the assigned label of the classified candidate unit.