Object Recognition Using Normal Maps from CAD Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recognition systems face challenges in accurately identifying objects from cluttered images, especially when only texture-less CAD models are available, due to the realism gap between synthetic and real data, requiring large amounts of labeled real images for training which is time-consuming and often impractical.
Innovation Solution
A method using artificial neural networks to generate normal maps from texture-less CAD models, allowing a task-specific recognition unit to identify objects directly from synthetic normal maps without the need for labeled real images, with an image generator trained to convert cluttered images into noiseless normal maps, enabling online processing and eliminating the need for extensive storage of synthetic images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If synthetic images are used for training the recognition system, then the need for large amounts of labeled real images is eliminated, but a realism gap exists between synthetic and real data reducing recognition accuracy
Solution Approach 1:
The patent introduces normal maps as an intermediary representation that bridges synthetic CAD models and real images. Normal maps encode surface geometry information in a format that can be generated from texture-less CAD models yet still capture essential object characteristics for recognition, serving as a mediator between the synthetic domain and real-world images.
Solution Approach 2:
The patent transforms the training approach by changing from using raw synthetic color images to using normal maps derived from CAD models. This parameter transformation allows the system to leverage geometric information from synthetic models in a way that better translates to real image recognition tasks, addressing the realism gap through representation transformation.
2Loss of information
If conventional GPU-based rendering is used to generate training images from CAD models, then color images with textures and lighting are produced, but the process is computationally intensive and requires extensive storage
Solution Approach 1:
The patent extracts only the essential geometric information from CAD models by generating normal maps rather than full-color rendered images. This extraction approach removes unnecessary visual details (textures, colors, lighting effects) while retaining the critical surface geometry needed for object recognition, significantly reducing computational and storage requirements.
Solution Approach 2:
Instead of generating realistic color images from CAD models and then trying to bridge the gap to real images, the patent inverts the approach by directly generating normal maps from CAD models that inherently capture geometric essence. This inverted approach works backwards from the goal of geometric understanding rather than visual realism.
Data Source
AI summary
The disclosure relates to a method how to recover an object from a cluttered image. The disclosure also relates to a computer program product and a computer-readable storage medium including instructions which, when the program is executed by a computer, cause the computer to carry out the acts of the mentioned method. Further, the disclosure relates to methods how to train components of a recognition system for recovering an object from such a cluttered image. In addition, the disclosure relates to such a recognition system.