Monocular Depth Estimation Using Histograms to Mask Dynamic-Object Errors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Monocular depth estimation models face challenges in accurately determining depth due to anomalies caused by dynamic objects, which result in aberrations such as infinite depth or incorrect depth values.
Innovation Solution
The proposed solution involves training a depth model using histograms to identify and mask erroneous loss values associated with anomalies in the image. This is achieved by generating histograms for different regions of the image based on photometric loss values, allowing the model to exclude these errors during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If self-supervised training is used to minimize photometric loss, then the depth model can be trained without ground truth depth data, but dynamic objects cause aberrations in depth values leading to reduced accuracy
Solution Approach 1:
The patent segments the image into multiple regions and generates separate histograms for each region. By dividing the image space into discrete regions, the system can identify and mask erroneous depth values caused by dynamic objects in specific regions without affecting the training process for other regions, thus maintaining training ease while improving depth estimation accuracy.
Solution Approach 2:
The patent introduces histograms as an intermediary mechanism between the photometric loss calculation and the training process. The histograms analyze the distribution of photometric loss values to identify anomalies caused by dynamic objects, allowing the system to filter out erroneous data before it corrupts the training, thereby resolving the contradiction between easy self-supervised training and accurate depth estimation.
2Device complexity
If photometric loss is applied to all image regions during training, then the training process is simple and unified, but erroneous values from dynamic objects corrupt the training signal
Solution Approach 1:
The training process is segmented into multiple steps: generating depth maps, calculating photometric loss, creating histograms for each region, identifying erroneous values through histogram analysis, and applying masks. This segmentation transforms a simple but unreliable unified process into a more complex but reliable multi-stage process that filters out corruption from dynamic objects.
Solution Approach 2:
The patent performs preliminary analysis of photometric loss values through histogram generation before the actual training update occurs. By identifying and masking erroneous values in advance, the system prevents corrupted data from affecting the training reliability, while maintaining a relatively simple overall framework.
3Quantity of substance
If depth values are estimated for all pixels, then complete depth coverage is achieved, but dynamic objects introduce infinite depth or incorrect depth values that degrade overall performance
Solution Approach 1:
The patent extracts and removes erroneous depth values from the training data by analyzing photometric loss histograms. Regions with dynamic objects that produce infinite or incorrect depth values are identified through abnormal histogram patterns and are masked out, allowing the system to maintain complete depth coverage for static regions while excluding only the problematic portions that degrade accuracy.
Data Source
AI summary
System, methods, and other embodiments described herein relate to an improved approach to training a depth model to derive depth estimates from monocular images using histograms to assess photometric losses. In one embodiment, a method includes determining loss values according to a photometric loss function. The loss values are associated with a depth map derived from an input image that is a monocular image. The method includes generating histograms for the loss values corresponding to different regions of a target image. The method includes, responsive to identifying erroneous values of the loss values, masking the erroneous values to avoid considering the erroneous values during training of the depth model.