3D-Controlled Synthetic Data Generation for Rare AI Scenarios
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models require large amounts of data, which are difficult and expensive to acquire, especially for rare situations, leading to insufficient or unsuitable training data, particularly in applications like fire and smoke detection at airports.
Innovation Solution
A method using a 3D engine to generate metadata for a reference scene, coupled with a control model and a generative model to produce synthetic images representative of the target situation, ensuring precise and faithful data generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real data is collected for training AI models, then the data is authentic and representative of real situations, but the data is difficult and expensive to produce and insufficient for rare situations
Solution Approach 1:
The patent creates synthetic copies of real-world scenes using 3D engines and generative models. Instead of collecting expensive real data, the system generates synthetic images that replicate the visual and semantic characteristics of real situations, including rare events. This copying approach maintains data authenticity while eliminating the high costs and logistical challenges of real data collection.
Solution Approach 2:
The system varies multiple parameters in the 3D engine (object positions, lighting conditions, camera angles, environmental factors) to generate diverse training scenarios. By systematically changing these parameters, the model produces a wide range of synthetic images that cover both common and rare situations, improving data diversity without additional field collection costs.
2Reliability
If real data is collected for rare situations, then the training data covers edge cases, but the quantity of data is insufficient
Solution Approach 1:
The system pre-generates large quantities of synthetic training data for rare situations before the AI model training begins. By using the 3D engine to create numerous variations of rare events (accidents, emergencies, edge cases) in advance, the system ensures that sufficient training examples are available, eliminating the bottleneck of insufficient rare event data.
Solution Approach 2:
The generative model creates multiple synthetic copies of rare situation templates. A single rare event scenario can be replicated hundreds or thousands of times with varying parameters (different locations, times, objects involved), exponentially increasing the quantity of rare situation data available for training without requiring proportional increases in real-world event collection.
3Adaptability or versatility
If a 3D engine is used to generate synthetic images, then full control over object positions and compositions is achieved, but the complexity of the system increases
Solution Approach 1:
The patent introduces a control model as an intermediary between the 3D engine and the generative model. This control model translates high-level scene descriptions and parameters into detailed 3D scene configurations, automatically managing the complexity of object placements, lighting, and camera settings. The intermediary layer simplifies the user interface while maintaining full control over scene parameters through automated parameter management.
4Ease of manufacture
If synthetic images are generated without control, then the generation process is simple, but the images are not realistic or faithful to the target situation
Solution Approach 1:
The control model serves as a mediator that guides the generative model to produce realistic images. It translates semantic constraints and scene parameters into controlled generation processes, ensuring that synthetic images maintain physical consistency, proper object relationships, and visual fidelity to real-world scenarios while keeping the generation process automated and relatively simple.
Solution Approach 2:
The system incorporates feedback mechanisms where the control model monitors and adjusts the generative process based on quality metrics and constraint satisfaction. This feedback loop ensures that generated images meet realism standards by automatically refining parameters and regenerating images that fail to meet quality thresholds, maintaining high reliability without manual intervention.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The invention relates to a method for generating data, the method being implemented by computer and comprising the steps: - implementing (22) a 3D engine to produce metadata relating to a reference scene generated by means of said 3D engine, the reference scene being representative of a predetermined target situation; - providing (24), as input to a control model coupled to a generative model, at least part of the metadata produced by the 3D engine; - calculating (26), by means of the generative model, at least one synthetic image representative of the target situation, from an output of the control model and descriptive data relating to the target situation; and - storing (34), in a data set, at least one calculated synthetic image.