Deformable Part Model 实现细节

本文介绍了一种基于特征金字塔和部分滤波器的视觉识别系统工作流程。系统通过计算滤波器响应、应用变形及结构规则,并使用动态规划来确定最优解。文章详细解释了符号与规则之间的对应关系及其在运行过程中的作用。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

Overview

Step 1. Compute filter responses to the feature pyramid

Each level of features (e.g. 46) is convolved with every stored part filters (e.g. 54), the result subsequently goes to the corresponding filter entry in model.symbols struct. Therefore each model.symbols{i}.score is a 1x46 cell.


Step 2. Apply deformation and structural rule

Deformation rule is applied first to each part filter (e.g. 54). New scores are saved to corresponding model.rules{i}.score, together with Ix and Iy


Structual rule is applied at last.

In current model, the number of structural rules equal to the number of component.

They are aggregated into one rule which is indicated by model.start.


Dynamic Prgramming is implemented during the realisation of structural rules


Step 3. Parse resulting scores to find optimal solution (can be multiple hypotheses)


Some Implementation Details

symbols <---> rules: 1-to1 correspondence

symbols 只有两种:terminal 和nonterminal

terminal symbol refers back to one entry in filters. In other words, terminal == filter


Run Time Variables

1. model.symbols{i}.score: 第i个filter对所有特征金字塔的响应。e.g. 1x46 cell。在filter_responses中被初次计算

2. model.rules{i}.[score | Ix | Iy]:应用第i条规则后的新的score值,以及位置信息。

3. model.scoretpt: TBD


function symbol_score

 - 对Deformatable Rule,只是简单的把新的score值放回symbols{i}.score中

 - 对Structure Rule,找出最大值(好像!)


评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值