Backbone Network Training via Masked Voxel Feature Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for inferring geometric properties from LiDAR data struggle to effectively improve the inference ability of neural networks, particularly by increasing the difficulty of inference through masking techniques.
Innovation Solution
A processor-implemented method that involves masking part of the voxel data from a point cloud to generate masked voxels, obtaining feature information from unmasked voxels through a backbone network, extracting prediction feature vectors, determining the parameter of a teacher module, and training the backbone network based on the similarity between prediction and masking feature vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If masking techniques are applied to increase inference difficulty, then the inference ability of neural networks is improved, but the training complexity and computational overhead increase
Solution Approach 1:
The teacher module is pre-trained before training the backbone network. This preliminary training allows the teacher module to provide accurate guidance features during the subsequent training process, improving the efficiency and effectiveness of the masking-based training without requiring complex iterative adjustments
Solution Approach 2:
A teacher module is introduced as an intermediary component that generates target feature vectors for masked voxels based on unmasked voxel information. This teacher module mediates the training process by providing supervised signals that guide the backbone network's learning, resolving the contradiction between increasing inference difficulty and maintaining training manageability
2Measurement precision
If more voxel data is masked to improve prediction accuracy, then the geometric property inference accuracy is improved, but the loss of information increases
Solution Approach 1:
The training process uses feedback from the teacher module's predictions on unmasked voxels to guide the backbone network's learning. The loss function computes the difference between the backbone network's predictions and the teacher module's predictions, providing directional feedback that enables accurate geometric property inference while maintaining information integrity through supervised learning signals
Data Source
AI summary
A processor-implement method includes masking at least a part of voxel data obtained from a point cloud to generate masked voxels, obtaining feature information about the masked voxels from unmasked voxels, which are not masked, through a backbone network, extracting a prediction feature vector from the feature information through a feature prediction model, determining a parameter of a teacher module based on a parameter of the backbone network, extracting a masking feature vector for the masked voxels through the teacher module, and training the backbone network by updating the parameter of the backbone network based on a similarity between the prediction feature vector and the masking feature vector.