Bounding-Box Template Matching for Hardware-Friendly Video Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The template matching process in video coding systems requires accessing external memory for reference data, leading to increased system bandwidth and computational complexity.
Innovation Solution
Implement a bandwidth-restricted template matching algorithm that restricts the search area to a bounding box based on the current motion vector, using techniques such as padding, pre-fetching, and sub-partitioning to reduce memory access and complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If template matching is performed using the full reference picture, then matching accuracy is improved, but memory bandwidth and computational complexity increase
Solution Approach 1:
The patent divides the reference picture into multiple regions based on motion vector prediction. Instead of using the entire reference picture for template matching, only the relevant region (first region) corresponding to the predicted motion is used. This segmentation reduces the amount of data accessed from external memory while maintaining matching accuracy for the current block.
Solution Approach 2:
The patent applies different processing strategies to different regions of the reference picture. The first region (corresponding to the current block's motion) is processed with full template matching, while other regions are either skipped or processed with reduced complexity. This local quality approach optimizes the balance between accuracy and computational resources.
2Measurement precision
If template matching search area is expanded, then matching accuracy is improved, but device complexity increases
Solution Approach 1:
The patent dynamically adjusts the template matching search area based on the motion vector prediction. The search area is restricted to a bounding box defined by the predicted motion vector, making the search region adaptive rather than fixed. This dynamic approach reduces computational complexity by limiting the search to relevant areas only.
Solution Approach 2:
The patent performs motion vector prediction first to determine the search region before performing template matching. This preliminary action allows the system to pre-calculate the bounding box and restrict the search area accordingly, reducing the computational complexity of the subsequent template matching process.
3Adaptability or versatility
If external memory is used for reference data, then reference picture storage is enabled, but system bandwidth increases
Solution Approach 1:
The patent extracts only the necessary portion of the reference picture (first region) from external memory for template matching, rather than loading the entire reference picture. This extraction approach reduces the data transfer volume from external memory, thereby reducing system bandwidth consumption while still enabling reference picture storage functionality.
Solution Approach 2:
The patent performs preliminary motion vector prediction and bounding box calculation before accessing external memory. This preliminary action determines exactly which region needs to be loaded from external memory, allowing the system to fetch only the necessary data and minimize bandwidth consumption during the reference picture storage and processing operation.
Data Source
AI summary
Method and apparatus for template matching with a determined area are disclosed. According to this method, a current template comprising current neighbouring pixels on an above side of the current block, on a left side of the current block, or a combination thereof for a current block is received. An area in a reference picture is then determined, where the reference picture corresponds to a previously coded picture. A matching result between a restricted reference template of a reference block and the current template is then determined, wherein the restricted reference template is generating by using only neighbouring reference pixels of a reference template inside the determined area, the reference template has a same shape as the current template, and a location of the reference template is determined according to a target motion vector (MV) from the current template.