Monocular Depth Estimation Aberration Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Monocular cameras struggle to accurately determine depth due to learned aberrations, particularly the infinite depth problem, where objects moving at the same speed as the camera are perceived as having infinite depth, leading to reduced situational awareness and navigation difficulties.
Innovation Solution
A two-stage training approach for a depth model, where the first stage uses self-supervised training with appearance-based loss to learn initial depth estimates, and the second stage filters out images causing aberrations by generating and analyzing depth maps to produce a subsampled training data set that avoids infinite depth issues, allowing for retraining of the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If self-supervised training with appearance-based loss is used to train the depth model, then the model can learn initial depth estimates from monocular images, but the model learns aberrations that cause objects moving at the same speed as the camera to be perceived as having infinite depth
Solution Approach 1:
The system performs preliminary action by training an initial depth model to identify aberrant images, then uses this model to filter the training data before final training. This two-stage approach prevents the infinite depth aberration from being learned in the first place, resolving the contradiction between initial learning capability and final prediction reliability
Solution Approach 2:
The system converts the harmful effect of appearance-based loss functions that cause infinite depth aberrations into a benefit by using the initially trained model to identify and filter out the problematic training images. The very mechanism that causes the aberration is leveraged to detect and remove the source of the problem, improving final depth estimation reliability
2Adaptability or versatility
If all training images are used including those with objects moving at the same speed as the camera, then the training data set is comprehensive, but the model learns aberrations that reduce situational awareness and navigation accuracy
Solution Approach 1:
The system extracts and removes the harmful subset of training images that cause infinite depth aberrations while retaining the beneficial diverse training data. By using the depth model to identify and filter out only the problematic images (those with objects moving at the same speed as the camera), the system maintains training data versatility while eliminating depth estimation errors
Solution Approach 2:
The system implements feedback by using the initially trained depth model to evaluate the training data quality and identify aberrant images. This feedback loop allows the system to automatically detect and remove problematic training samples, ensuring that the final training process produces reliable depth estimates without sacrificing data diversity
Data Source
AI summary
System, methods, and other embodiments described herein relate to self-supervised training for monocular depth estimation. In one embodiment, a method includes filtering disfavored images from first training data to produce second training data that is a subsampled version of the first training data. The disfavored images correspond with anomaly maps within a set of depth maps. The first depth model is trained according to the first training data and generates the depth maps from the first training data after initially being trained with the first training data. The method includes training a second depth model according to a self-supervised training process using the second training data. The method includes providing the second depth model to infer distances from monocular images.