One-stage Object Detection Training with Scale-specific Focal Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing one-stage object detecting models face poor training effects due to the use of a unified loss function for feature maps at different scales, which ignores scale differences and reduces model performance.
Innovation Solution
The method involves acquiring and determining separate loss function values for each scale during training, using a focal loss function to optimize feature maps hierarchically, thereby improving model performance without adding extra parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a unified loss function value is used for feature maps at different scales, then the model structure is simple, but the training effect is poor and model performance is reduced
Solution Approach 1:
The patent applies local quality by assigning different loss function values to feature maps at different scales. Instead of using a unified loss value, the system calculates separate loss values for each scale (e.g., P2, P3, P4, P5) based on the specific characteristics and requirements of each scale's feature maps, thereby optimizing training effectiveness for each scale independently
Solution Approach 2:
The patent changes the parameter of loss function values from a unified constant to scale-dependent variables. By introducing scale-specific loss values (loss_P2, loss_P3, loss_P4, loss_P5) that vary according to the scale of feature maps, the system enables differentiated training optimization without fundamentally changing the overall model architecture
2Productivity
If a unified loss function value is used for all scales, then the training process is simple, but model performance is reduced due to ignoring scale differences
Solution Approach 1:
The system implements local quality by treating different scales with different loss weights. Each scale's feature maps receive customized loss function values that reflect their specific contributions to detection accuracy, allowing the training process to optimize for each scale's unique characteristics rather than applying a one-size-fits-all approach
Solution Approach 2:
The patent introduces dynamics by making the loss function values adaptive rather than static. The loss values for different scales are dynamically adjusted during training based on the actual performance and characteristics of each scale, enabling the system to automatically optimize training efficiency and performance without manual intervention
Data Source
AI summary
Technical solutions relate to the field of artificial intelligence such as deep learning, computer vision and intelligent imaging. A method may includes during training of a one-stage object detecting model, acquiring values of a loss function corresponding to feature maps at different scales respectively in the case that classification loss calculation is required, and the loss function is a focal loss function; and determining a final value of the loss function according to the acquired values of the loss function, and training the one-stage object detecting model according to the final value of the loss function.