Texture Generation Removing Moving Vehicles from 3D Meshes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for texturing 3D meshes of urban areas often result in crowded and visually unpleasing outcomes due to the inclusion of moving vehicles, leading to artefacts like ghosting and incomplete sampling of objects.
Innovation Solution
A computer-implemented method that uses semantic segmentation and structure-from-motion algorithms to identify and remove moving vehicles from input images, replacing them with authentic or plausible background pixels to generate a clean texture for 3D meshes, thereby preventing artefacts and enhancing visual appeal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If moving vehicles are included in the texture generation from input images, then the texture comprehensively represents all objects in the scene, but the texture becomes crowded and visually unpleasing with artefacts
Solution Approach 1:
The patent extracts and removes moving vehicle representations from the texture generation process. By identifying vehicles as distinct target objects and excluding them from the final texture, the method eliminates the harmful visual artefacts and crowding effects while preserving the static scene elements that provide comprehensive scene representation.
Solution Approach 2:
The patent applies dynamic detection to identify moving vehicles across multiple input images. By detecting objects that change position between images, the system dynamically separates moving vehicles from static scene elements, allowing selective exclusion of vehicles from the texture while maintaining representation of stationary objects.
2Productivity
If direct comparison of multiple images is used to detect moving objects, then the detection process is simple and fast, but the method cannot reliably distinguish moving vehicles from static objects in complex urban scenes
Solution Approach 1:
The patent performs preliminary semantic segmentation on each input image before comparing them to detect moving objects. By pre-classifying objects into categories such as vehicles, buildings, and vegetation, the system establishes a foundation for accurate moving object detection that goes beyond simple pixel comparison, improving classification accuracy while maintaining efficiency.
Solution Approach 2:
The patent uses feedback from semantic segmentation results to guide the moving object detection process. The segmentation labels provide contextual information that helps distinguish true moving vehicles from static objects that may appear to move due to perspective or occlusion, thereby improving detection precision without significantly increasing computational complexity.
Data Source
AI summary
A method for creating a texture from input images, thereby removing representations of objects from the texture, the method comprising performing semantic segmentation in a plurality of digital input images with a plurality of semantic classes, at least one of the semantic classes relating to a target object class; identifying, in at least a first input image, one or more possible instances of representations of objects belonging to the target object class, each possible instance being constituted by a plurality of contiguous instance pixels of the image pixels; determining whether the instance pixels are target pixels, target pixels being pixels constituting a representation of an object belonging to the target object class; and replacing target pixels in the texture with replacement pixels derived from one or more of the plurality of digital input images.