Dynamic Proposal Object Detection for Resource-Constrained Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection algorithms in computer vision systems use a fixed number of proposals, which can be inefficient on devices with resource constraints and wasteful when image complexity is low, as they cannot adaptively adjust the number of proposals based on image complexity or available resources.
Innovation Solution
A dynamic or switchable number of proposals is introduced, where the number of proposals can be determined based on conditions such as image complexity and resource availability, allowing the algorithm to adaptively adjust its computational resources, and an in-place distillation training process is used to improve model variants with fewer proposals by transferring knowledge from models with more proposals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a fixed number of proposals is used in object detection algorithms, then the algorithm structure is simple and easy to implement, but the computational efficiency deteriorates on devices with resource constraints and wastes resources when image complexity is low
Solution Approach 1:
The patent applies dynamics by transitioning from a fixed number of proposals to a dynamic number of proposals that can be adjusted based on image complexity and device resources. The system automatically determines the appropriate number of proposals (e.g., 96, 192, or 288) based on the input image characteristics and available computational resources, optimizing the balance between accuracy and efficiency without requiring manual intervention or complex model switching mechanisms.
2Adaptability or versatility
If a fixed number of proposals is used in object detection algorithms, then the algorithm is straightforward to deploy, but computational resources are wasted when image complexity is low
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the number of proposals as a key parameter based on image complexity metrics and device resource availability. Instead of using a constant number of proposals throughout all scenarios, the system changes this critical parameter adaptively - using fewer proposals (e.g., 96) for simple images and more proposals (e.g., 288) for complex images - thereby optimizing computational resource utilization across diverse imaging conditions.
3Productivity
If a dynamic number of proposals is used, then computational efficiency and adaptability improve, but the algorithm complexity and difficulty of implementation increase
Solution Approach 1:
The patent applies self-service by implementing an automatic mechanism where the system itself determines the optimal number of proposals based on image complexity analysis and device resource detection, without requiring external intervention or complex user configuration. The algorithm self-adjusts the proposal count based on pre-computed image characteristics and available computational power, simplifying deployment while maintaining high adaptability and efficiency.
Data Source
AI summary
Described are examples for detecting objects in an image on a device including setting, based on a condition, a number of sparse proposals to use in performing object detection in the image, performing object detection in the image based on providing the sparse proposals as input to an object detection process to infer object location and classification of one or more objects in the image, and indicating, to an application and based on an output of the object detection process, the object location and classification of the one or more objects.