Tile-Based Non-Maximum Suppression for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection processes, such as single shot detection, often generate multiple detection boxes for a single object, leading to complexity in suppressing redundant boxes, and current non-maximum suppression (NMS) methods like IoU-based approaches can be computationally intensive and inefficient.
Innovation Solution
A tile-by-tile suppression mechanism is proposed, where an image is divided into tiles, and detection boxes overlapping within each tile are compared to determine which boxes to discard based on confidence values, reducing the number of comparisons and processing steps through parallelization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional non-maximum suppression (NMS) is performed by comparing each detection box with all other detection boxes, then suppression accuracy is improved, but computational complexity increases significantly
Solution Approach 1:
The image is divided into multiple tiles, and detection boxes are grouped by the tiles they overlap. This segmentation reduces the comparison scope from all detection boxes to only those within the same tile, significantly reducing computational complexity while maintaining suppression accuracy for local overlapping boxes.
Solution Approach 2:
The patent performs NMS only on detection boxes that overlap the same tile, rather than comparing all detection boxes globally. This partial action approach reduces computational burden while still achieving effective suppression for boxes that actually overlap, as non-overlapping boxes don't require suppression.
2Reliability
If detection boxes are sorted by confidence before performing NMS, then suppression reliability is improved, but processing time increases
Solution Approach 1:
The patent sorts detection boxes by confidence separately within each tile group rather than sorting all detection boxes globally. This segmented sorting approach maintains suppression reliability within local overlapping regions while reducing the time required for sorting operations.
Solution Approach 2:
Confidence-based sorting is applied only to detection boxes within the same tile group, not to all detection boxes across the entire image. This partial sorting maintains reliability for suppressing overlapping boxes while significantly reducing processing time compared to global sorting.
3Productivity
If the number of detection boxes is reduced through aggressive suppression, then processing efficiency is improved, but detection accuracy may deteriorate
Solution Approach 1:
The patent applies suppression with different criteria based on local conditions - only suppressing detection boxes that overlap the same tile. This local quality approach ensures that suppression is applied appropriately only where needed (overlapping boxes) while preserving non-overlapping boxes, maintaining detection accuracy while improving processing efficiency.
Solution Approach 2:
The patent replaces the traditional global NMS mechanical process with a tile-based grouping and comparison mechanism. This substitution allows for more efficient processing by organizing detection boxes into manageable tile groups before comparison, improving productivity without sacrificing detection accuracy.
Data Source
AI summary
A mechanism for performing non-maximum suppression (NMS) on a plurality of detection boxes identifying potential locations for one or more objects within an image. The mechanism uses a tiling system that divides the image into a plurality of tiles. A tile-by-tile suppression process is performed, in which at least some detection boxes that overlap a particular tile are processed to determine whether any detection boxes are to be discarded.