Regression Focal Loss for Optical-Flow Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in accurately training machine-learning models for optical-flow and disparity-determination techniques due to factors like feature saliency, lighting conditions, and occlusions, which affect pixel-level correspondence across images.
Innovation Solution
The use of a Regression Focal Loss (RFL) that weights predictions based on inverse confidence for each pixel, combined with confidence maps to guide training losses, allowing the model to focus on difficult samples and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard loss functions are used for training optical-flow models, then training process is simple, but model accuracy deteriorates due to inability to handle difficult samples
Solution Approach 1:
The patent applies local quality by creating a confidence map that assigns different weights to different spatial locations and pixels based on their reliability. High-confidence regions are weighted more heavily in the loss calculation, while low-confidence regions (such as those with occlusions or low feature saliency) are downweighted. This allows the model to focus training attention on the most informative regions, improving overall accuracy without requiring complete redesign of the training framework.
Solution Approach 2:
The patent changes the loss function parameters dynamically by introducing a confidence map as an additional parameter that modulates the loss weights. Instead of using uniform loss weights, the system computes confidence values for each pixel based on feature saliency and occlusion detection, then uses these confidence values to adjust the loss function parameters during training. This adaptive parameter adjustment enables the model to handle difficult samples effectively.
2Reliability
If all pixels are treated equally in loss calculation, then computation is efficient, but performance on difficult samples (occlusions, low saliency) deteriorates
Solution Approach 1:
The patent segments the image into different regions based on confidence levels by detecting feature saliency and occlusions. The confidence map divides pixels into high-confidence and low-confidence groups, allowing the loss function to apply different treatment to each segment. This segmentation enables the model to focus computational resources on training difficult samples without completely abandoning efficient processing of easy samples.
Solution Approach 2:
The patent applies partial action by selectively applying higher loss weights only to difficult samples and low-confidence regions, rather than uniformly increasing all loss weights. This partial focus on difficult samples (occlusions, low saliency areas) allows the model to improve reliability in these specific regions while maintaining overall training efficiency by not over-emphasizing all pixels equally.
3Measurement precision
If confidence maps are computed for every pixel, then training accuracy improves, but computational cost increases
Solution Approach 1:
The patent performs preliminary action by pre-computing the confidence map during the training process rather than requiring it for every inference. The confidence map is generated once during training based on feature saliency and occlusion detection, and then used to guide the loss calculation. This preliminary computation allows the system to achieve high pixel-level correspondence accuracy without requiring expensive real-time confidence map generation during actual model inference.
Data Source
AI summary
Systems and techniques are described herein for processing image data. For instance, a method for processing image data is provided. The method may include predicting, using a machine-learning model, a difference map indicative of differences between a first image and a second image to generate a predicted difference map; determining a confidence map based on the predicted difference map; determining an error based on the confidence map and a comparison of the predicted difference map and a ground-truth difference map; and adjusting one or more parameters of the machine-learning model based on the error.