Self-Supervised Depth and Pose Estimation With Pose Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing depth and pose estimation methods in autonomous systems, such as LIDAR and stereo cameras, suffer from inaccurate pose estimates and difficulty in computing ground truth depth, especially in dynamic real-world scenes, leading to suboptimal performance in tasks like obstacle avoidance and mapping.
Innovation Solution
A method employing a self-supervised deep learning approach that generates additional pose changes using a Gaussian distribution to improve camera pose and depth map estimation, leveraging neural networks for enhanced accuracy through sampling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If deep learning models are used for depth estimation, then the system can operate without ground truth depth, but the pose estimate accuracy deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the initially estimated pose change is used to generate additional pose changes, which are then evaluated by reconstructing images and comparing them with the target image. The best matching pose change is selected based on this feedback, thereby improving pose estimation accuracy while maintaining the ability to operate without ground truth depth.
Solution Approach 2:
The patent performs a preliminary pose change estimation using deep learning models, then uses this preliminary result to generate additional candidate pose changes. This preliminary action provides a starting point that guides the subsequent refinement process, enabling the system to operate without ground truth while still achieving accurate pose estimates.
2Device complexity
If independent depth and pose estimation are performed with separate loss functions, then computation is simplified, but estimation accuracy deteriorates
Solution Approach 1:
The patent merges the independent depth and pose estimation processes by using the depth map as a shared resource between both estimations. The pose change estimation leverages the depth map to generate additional candidate pose changes, and both estimations are evaluated together through image reconstruction and comparison, thereby improving accuracy while maintaining computational simplicity.
3Measurement precision
If additional pose changes are generated through sampling, then pose estimation accuracy improves, but computation time increases
Solution Approach 1:
The patent performs a preliminary pose change estimation using deep learning models to obtain an initial result. This preliminary action provides a centered value around which additional pose changes are sampled, reducing the search space and computation time compared to sampling from scratch, while still achieving improved accuracy through the evaluation of multiple candidates.
Solution Approach 2:
The patent changes the parameter space by generating additional pose changes through sampling around the initial estimate rather than exhaustively searching all possible pose changes. This parameter change strategy focuses computation on the most likely candidates, improving accuracy while limiting computation time.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Provided is a method for a camera pose change estimation. A first image (106) of a scene (102, 202) is captured with a first camera (104, 304). A depth map is obtained with respect to the first camera (104, 304) based on the first image (106). A second image (110) of the scene (102, 202) is captured with a second camera (108, 306). A pose change is obtained from the first camera pose to the second camera pose based on the first image (106) and the second image (110). A set of additional pose changes is generated based on a pose change. A set of reconstructed images is obtained and each of the reconstructed images is matched with the second image (110). A camera pose change estimation is selected among the pose change and the set of additional pose changes that corresponds to a best matching reconstructed image.