Direct Depth Prediction via Non-Linear Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for depth prediction in multi-camera setups face challenges due to the lack of correlation between disparity data and depth data, leading to inaccurate depth estimation and scale ambiguity.
Innovation Solution
The use of a non-linear mapping function applied to the output of a layer of a machine learning model to directly generate depth data, allowing for more accurate and reliable depth prediction without the need for disparity conversion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning models use disparity data for depth prediction in multi-camera setups, then the model can process image data from multiple cameras, but the depth estimation accuracy deteriorates due to lack of correlation between disparity data and depth data
Solution Approach 1:
The patent extracts and removes the disparity data conversion step from the depth prediction pipeline. Instead of converting disparity data to depth data through conventional methods, the model directly predicts depth values from image data, eliminating the source of correlation loss and improving depth estimation accuracy
Solution Approach 2:
The patent introduces a non-linear mapping function as an intermediary between the machine learning model output and the final depth prediction. This mapping function (e.g., exponential or logarithmic) transforms the model's raw output into accurate depth values, serving as a mediator that resolves the correlation issue between disparity and depth data
2Measurement precision
If conventional methods convert disparity data to depth data, then depth prediction can be performed, but scale ambiguity increases leading to inaccurate depth estimation
Solution Approach 1:
The patent removes the disparity-to-depth conversion process entirely, replacing it with direct depth prediction. This extraction of the problematic conversion step eliminates the scale ambiguity that arises from conventional disparity-based methods, as the model learns depth directly without relying on disparity conversions
Solution Approach 2:
The patent changes the output parameter of the machine learning model from disparity values to direct depth values. By modifying what the model predicts (from disparity to depth) and applying non-linear mapping functions to the output, the scale ambiguity inherent in disparity-based approaches is resolved
Data Source
AI summary
Example systems and techniques are described for training a machine learning model. A system includes memory configured to store image data captured by a plurality of cameras and one or more processors communicatively coupled to the memory. The one or more processors are configured to execute a machine learning model on the image data, the machine learning model including a plurality of layers. The one or more processors are configured to apply a non-linear mapping function to output of one layer of the plurality of layers to generate depth data. The one or more processors are configured to train the machine learning model based on the depth data to generate a trained machine learning model.