Scene Flow Label Generation Using Object Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for predicting scene flow in autonomous vehicles face challenges in training neural networks due to the lack of accurate correspondence between points in different point clouds, making it difficult to generate ground truth motion vectors, and are computationally inefficient for large point clouds.
Innovation Solution
The system generates scene flow labels using object label data to identify previous positions of points, allowing for accurate estimation of motion vectors, and employs a neural network architecture optimized for low-latency processing of large point clouds, enabling efficient deployment on autonomous vehicles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing techniques are used to predict scene flow directly from point clouds, then the prediction can be made, but accurate ground truth motion vectors cannot be generated due to lack of point correspondence information
Solution Approach 1:
The patent introduces object label data as an intermediary to establish point correspondence between different point clouds. Instead of directly matching points between time points, the system uses object labels to identify which points belong to the same objects, thereby inferring correspondence relationships. This mediator enables accurate ground truth motion vector generation by providing the missing linkage information between temporal point clouds.
2Productivity
If neural networks are trained to predict scene flow from point clouds, then motion prediction can be achieved, but the computational efficiency is insufficient for large point clouds
Solution Approach 1:
The patent segments the point cloud processing by dividing it into object-level components. Instead of processing all points uniformly, the system groups points into objects based on object label data, allowing the neural network to process smaller, more manageable object instances separately. This segmentation reduces the computational burden while maintaining prediction accuracy for large point clouds.
3Quantity of substance
If point clouds from laser sensors are used, then scene data can be captured, but no information about correspondence between points in different time point clouds is available
Solution Approach 1:
The patent applies preliminary action by generating object label data in advance, before the actual scene flow prediction task. These pre-computed object labels serve as a reference framework that captures correspondence relationships between points across different time points. By preparing this correspondence information beforehand, the system eliminates the need for complex real-time point matching during prediction.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for predicting scene flow. One of the methods includes obtaining a current point cloud representing an observed scene at a current time point; obtaining object label data that identifies a first three-dimensional region in the observed scene; determining, for each current three-dimensional point that is within the first three-dimensional region and using the object label data, a respective preceding position of the current three-dimensional point at a preceding time point in a reference frame of the sensor at the current time point; and generating, using the preceding positions, a scene flow label for the current point cloud that comprises a respective ground truth motion vector for each of a plurality of the current three-dimensional points.