Bird's-Eye View Image Synthesis With Artifact Removal and Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing bird's eye view (BEV) image generation techniques often produce distorted and artifact-laden images due to transformation processes, leading to inaccurate and difficult-to-process results, which negatively impact driver assistance and automation features.
Innovation Solution
A system utilizing a machine learning component, comprising a generator and discriminator, to produce synthesized BEV images and segmentation maps by removing artifacts from stitched images, employing generative adversarial models and shared encoding/decoding layers to optimize image generation and prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If transformation processes are used to generate BEV images, then BEV images can be produced for driver assistance functions, but distortion and artifacts are introduced that reduce accuracy
Solution Approach 1:
The patent introduces an intermediary processing step between the transformation process and the final BEV image output. A correction module or post-processing algorithm acts as a mediator to identify and remove artifacts and distortion introduced during the transformation, thereby recovering image accuracy while maintaining the productivity benefits of transformation-based generation.
Solution Approach 2:
The patent replaces the traditional mechanical/geometric transformation approach with a learning-based system. Instead of relying solely on geometric transformations that inherently introduce distortion, the system uses trained models (such as neural networks) to learn the correct mapping from input images to BEV images, substituting the rigid mechanical transformation with a flexible, data-driven approach that maintains accuracy.
2Manufacturing precision
If artifact removal is applied to improve BEV image accuracy, then image quality improves, but processing complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training artifact removal models on synthetic data that includes various types of transformation artifacts. This pre-processing of training data and pre-training of models allows the system to efficiently remove artifacts during actual BEV generation without adding significant runtime complexity, as the heavy lifting of learning artifact patterns is done beforehand.
Solution Approach 2:
The patent uses copying by creating synthetic training examples that replicate real-world transformation artifacts. By generating copies of artifacts through simulated transformations and using these copies to train removal algorithms, the system can learn to recognize and remove actual artifacts efficiently without requiring complex real-time analysis of each artifact instance.
Data Source
Figure 1A~1C
Figure 2
Figure 3
AI summary
Techniques are described for generating bird's eye view (BEV) images and segmentation maps. According to one or more embodiments, a system is provided comprising a processor that executes computer executable components stored in at least one memory, comprising a machine learning component that generates a synthesized bird's eye view image from a stitched image based on removing artifacts from the stitched image present from a transformation process. The system further comprising a generator that produces the synthesized bird's eye view image and a segmentation map, and a discriminator that predicts whether the synthesized bird's eye view image and the segmentation map are real or generated.