Depth Estimation Model Training With Selective Metric Depth Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing depth estimation models struggle to achieve ideal results in various subdivided scenes, despite advancements in self-supervision and relative depth learning, necessitating improved generalization and accuracy in both indoor and outdoor environments.
Innovation Solution
A training method for depth estimation models involving selective depth data acquisition, including quality assessment, mean-shift clustering, fine-tuning, and joint supervision with gradient angle functions to enhance model performance in complex scenes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large amount of data is used for pre-training to improve generalization ability in coarse scenes, then the model achieves better performance in indoor and outdoor scenes, but the model still struggles to achieve ideal results in various subdivided scenes
Solution Approach 1:
The patent segments the training data into different quality levels and applies different processing strategies. High-quality depth data is selected through quality assessment and mean-shift clustering to form a specialized training subset, while general data is used for pre-training. This segmentation allows the model to learn both general patterns and fine-grained details separately, resolving the contradiction between generalization and subdivided scene accuracy.
Solution Approach 2:
The patent applies local quality by treating different data subsets with different quality standards. High-quality data undergoes rigorous selection and clustering processes, while the overall training process maintains diversity. The model is trained with focused attention on high-quality data for specific scene types, allowing different parts of the training process to have different quality characteristics, thus improving both generalization and subdivided scene performance.
2Measurement precision
If traditional viewing angle matching or triangulation methods are used, then depth estimation can be obtained, but the method lacks the generalization ability and accuracy achieved by self-supervision and relative depth learning
Solution Approach 1:
The patent employs self-supervision learning where the model learns to estimate depth by supervising itself on datasets with known depth information. The model predicts depth maps and compares them against ground truth or relative depth constraints, enabling it to learn depth estimation patterns without requiring complex external calibration or triangulation setups. This self-service approach improves both accuracy and generalization across diverse scenes.
Solution Approach 2:
The patent performs pre-training on large-scale datasets with relative depth information before fine-tuning on high-quality metric depth data. This preliminary action allows the model to learn general depth estimation patterns and scene understanding first, then refine its accuracy on specific high-quality data, achieving both generalization and precision.
3Measurement precision
If mean-shift clustering is applied to select seed depth data, then the model can focus on high-quality data, but the training process becomes more complex
Solution Approach 1:
The patent extracts high-quality depth data from the overall dataset through quality assessment and mean-shift clustering. By separating and focusing on the most valuable data subset (seed depth data), the model can achieve higher accuracy without needing to process the entire large dataset in detail. This extraction approach improves precision while managing training complexity through selective processing.
Solution Approach 2:
The patent applies partial action by using mean-shift clustering on a representative subset of data to identify seed depth data, rather than applying complex processing to the entire dataset. This partial processing of high-value data provides sufficient accuracy improvement without incurring the full complexity cost of exhaustive analysis, achieving a practical balance between precision and complexity.
Data Source
AI summary
A training method of depth estimation model, terminal and storage medium are provided by the present disclosure. The method includes: acquiring selective depth data; training the depth estimation model by using the selective depth data to obtain a trained depth estimation model; where the acquiring selective depth data includes: acquiring a depth data set; performing quality assessment on the depth data set to obtain first depth data; performing mean-shift on the first depth data to obtain second depth data; performing fine-tuning on a pre-training depth model by using the second depth data to obtain a metric depth model; and performing necessity assessment on the first depth data by using the metric depth model to obtain the selective depth data.