Depth Estimation Model Training for Weak-Texture Image Regions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing depth estimation methods struggle to accurately estimate depth in low-texture images due to large errors, leading to reconstructed frames that resemble the target frame despite significant depth inaccuracies.
Innovation Solution
A method involving a depth estimation model trained using autoencoders, posture estimation, and texture-based region weighting to adjust parameters based on similarity calculations, reducing the impact of weak texture regions during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If monocular depth estimation is used with structure from motion, then depth values can be estimated and images reconstructed, but the similarity of reconstructed frames cannot accurately express error degree in low texture scenarios
Solution Approach 1:
The patent applies local quality by differentiating loss calculation across different image regions. Specifically, it identifies weak texture regions and applies different loss weights to these regions versus non-weak texture regions. This allows the model to focus more on regions where depth estimation is more reliable and less on regions where similarity metrics are misleading, thereby resolving the contradiction between depth estimation and error detection capability.
2Ease of manufacture
If standard similarity-based loss function is used for training, then model training is simplified, but weak texture regions cause large depth errors to be learned as correct
Solution Approach 1:
The patent introduces local quality by computing a texture strength map and using it to weight the loss function differently across regions. Weak texture regions are assigned lower weights, preventing them from dominating the training process. This maintains relative training simplicity while dramatically improving depth estimation precision in challenging regions.
Solution Approach 2:
The patent changes the loss function parameters by introducing region-dependent weights. Instead of a uniform loss function, it dynamically adjusts the contribution of different image regions to the total loss based on their texture characteristics. This parameter change allows the model to learn more accurate depth representations without completely redesigning the training framework.
3Power
If uniform loss weighting is applied across all regions, then training computation is reduced, but depth estimation accuracy deteriorates in weak texture regions
Solution Approach 1:
The patent implements local quality by computing texture strength for different regions and applying localized loss weighting. This approach maintains computational efficiency by using simple texture analysis (gradient-based) while significantly improving depth estimation accuracy in weak texture regions through intelligent loss distribution.
Solution Approach 2:
The patent applies partial action by focusing computational effort on non-weak texture regions where depth estimation is more reliable, while applying reduced weighting to weak texture regions. This selective approach improves overall accuracy without requiring excessive computational resources to process all regions equally.
Data Source
AI summary
A method for training a depth estimation model comprise acquires a first image and a second image being inputted into the depth estimation model. The depth estimation model outputs a first depth image. A posture conversion relationship between the first image and the second image is extracted by a posture estimation model. A restored image is generated based on the first depth image, the posture conversion relationship, and pre-obtained camera parameters. A similarity between the restored image and the first image is calculated to obtain a two-dimension loss image. A first similarity of pixel points of each weak texture region are determined based on the two-dimension loss image. A ratio of the first similarity for adjusting the parameters of the depth estimation model is decreased and a first loss value is obtained. A training apparatus and an electronic device applying the method are also disclosed.