Object Detection via Temporal Bounding Box Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accurately and efficiently reconstructing the motion of the human hand from images remains challenging due to unconstrained global and local pose variations, frequent occlusion, local self-similarity, and high articulation.
Innovation Solution
An object detection method and system that involves obtaining candidate bounding boxes of an interest object in a current image frame and filtering them based on a determined bounding box from a previous image frame, using a machine-learning model with a backbone feature extractor, bounding box detection head, and bounding box selection head.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional object detection methods are used to detect hand motion from images, then the system can process unconstrained pose variations and occlusions, but the detection accuracy and efficiency deteriorate due to local self-similarity and high articulation
Solution Approach 1:
The patent applies preliminary action by using the determined bounding box from the previous image frame to pre-filter candidate bounding boxes in the current frame before final detection. This constrains the search space in advance, improving both accuracy and efficiency by eliminating obviously incorrect candidates early in the process.
Solution Approach 2:
The patent implements feedback by using the determined bounding box from the previous frame as a reference to filter and validate candidate bounding boxes in the current frame. This temporal feedback mechanism leverages motion continuity to improve detection accuracy while maintaining efficiency through intelligent candidate elimination.
2Reliability
If candidate bounding boxes are generated without constraints from previous frames, then the system maintains flexibility for unconstrained pose variations, but the robustness of detection deteriorates due to frequent occlusion and high articulation
Solution Approach 1:
The patent uses the determined bounding box from the previous frame to pre-filter candidate bounding boxes before final detection. This preliminary constraint improves robustness by eliminating candidates that are unlikely to be correct, while still allowing the system to adapt to pose variations through the candidate generation process.
Solution Approach 2:
The patent applies dynamics by making the filtering constraints adaptive rather than fixed. The determined bounding box from the previous frame serves as a dynamic reference that adjusts to motion while maintaining temporal consistency, allowing the system to handle both robustness requirements and pose flexibility.
3Productivity
If the system processes all candidate bounding boxes without filtering, then it maintains completeness for high articulation objects, but the computational complexity increases
Solution Approach 1:
The patent extracts and utilizes the determined bounding box from the previous frame as a filtering criterion to eliminate unnecessary candidate bounding boxes from further processing. This extraction of temporal information reduces computational complexity by focusing resources on promising candidates only.
Solution Approach 2:
The patent performs preliminary filtering of candidate bounding boxes using the previous frame's determined bounding box before conducting detailed detection. This preliminary action reduces the number of candidates requiring full processing, thereby improving productivity while maintaining detection quality for high articulation objects.
Data Source
AI summary
The embodiments of the present disclosure disclose an object detection method and system, and a non-transitory computer-readable medium. In the object detection method, multiple candidate bounding boxes of an interest object in a current image frame are obtained. Based on a determined bounding box of the interest object in a previous image frame, the multiple candidate bounding boxes are filtered to obtain a determined bounding box of the interest object in the current image frame.