BIM(biologically inspired model)小结

BIM(biologically inspired model)小结

1. 模型相关

1.1 Thomas Serre(MIT)

Publication:

Object recognition with features inspired by visual cortex (CVPR2005)

Robust Object Recognition with Cortex-Like Mechanisms (PAMI-07)

内容:主要讲C2 features

S1(Gabor) ——C1(local max)——S2(computing, learning)——C2(global max)

实验:检测物体是否存在。

DataSets;MIT_CBCL datasets , Caltech-101

Linear SVM + gentle adaboost

Object detection:in clutter, without clutter, of texture based objects(buildings ,trees, roads, skies)

Scene understanding: Objects分为:texture-based和shape-based。

源码:

可下载,已跑通,比较熟悉。

1.2 Jim Mutch(MIT):作者是Serre的师弟,将BIM用于物体识别和定位。

Publication:

Multiclass Object Recognition with Sparse, Localized Features(CVPR 06)

Object class recognition and localization using sparse features with limited receptive fields(IJCV 08)

内容:对Serre的BIM模型提出了改进,主要包括

Sparsify S2 inputs.(选主方向)

Inhibit S1/C1 outputs. (设阈值)

Limit position/scale invariance of S2 features.

Select features that are highly weighted by the SVM.

实验:多类物体分类问题

(1) Multiclass 实验(caltech 101),主要讲参数的选择。

(2) 检测和定位:features(BIF), 滑动窗算法(PAMI 04)。

(3) 定位实验:Graz-02(IJCV)。

源码:可下载,已跑通。

1.3 Yongzheng Huang(Institute of Automation):EBIM

Publication:

Enhanced Biologically Inspired Model (CVPR2008)

Enhanced Biologically Inspired Model(SMC 11)

内容:针对BIM的缺点(a very heavy computational cost due to dense input, a disputable pooling operation in modeling relations of the visual cortex, and blind feature selection in a feed forward framework.)进行了改进,主要有

(1) sparsity constraints

(2) local weighted sum(SMC 中将这部分去掉了)

BIM(biologically inspired model)小结 - 淮静 - 淮静的博客

(3) feedback: boosting+ SVM

实验:物体识别和检测

Datasets: Caltech 5, MIT-CBCL street scene, the GRAZ , the PASCAL 07

SMC中Discussion部分分析的很好,对Serre BIM,Mutch BIM和EBIM做了对比。EBIM用梯度去了噪声,再用boosting+svm的方法减少Patch的数量。而不是像Mutch那样去除了一些信息。

2. 应用相关

2.1 Guodong Guo, Yun Fu (UIUC)

Publication:

Human Age Estimation Using Bio-inspired Features(CVPR 09),重点读了一下

A Study on Automatic Age Estimation using a Large Database(ICCV 09)

CVPR的内容:

C1 + feature dimension reduction + statistical learning

主要改进之处,在文章的” Studies of Improvement”部分讲的很清楚。

C1 vs. C2 features

The number of bands and orientations

Filter banks starting from 5 × 5 vs. 7 × 7

“STD” vs. “MAX”

BIM(biologically inspired model)小结 - 淮静 - 淮静的博客

Dimensionality reduction by PCA

2.2 Ethan Meyers(MIT)

Publication: Using Biologically Inspired Features for Face Processing(IJCV 08)

内容:

(1) 在S1前:center-surround

(2) 由C1到S1时进行了线性加权(KR-RCA)

2.3 Dacheng Tao (University of Technology, Sydney)

2.3.1 Yang Mu

Publication:

Biologically Inspired Tensor Features (Cogn Comput 09)

Biologically inspired feature manifold for gait recognition(Neurocomputing 10)

Biologically Inspired Model for Crater Detection (IJCNN 11) 重点读了一下

IJCNN 11主要思路:

C1(location, scale invariance) + Harr (rotation invariance) + 降维(Discriminative locality alignment ,DLA)

实验:Crater detection等

2.3.2 Dongjin Song

Publication:

C1 Units for Scene Classification (ICPR 08)

DISCRMINATIVE GEOMETRY PRESERVING PROJECTIONS (ICIP 09)

Biologically Inspired Feature Manifold for Scene Classification (TIP 10) 重点读了一下

TIP 主要内容:

将BIF用于场景分类,由三部分组成:

1) BIF: intensity,color(itti的方法),C1 features

2) DGPP:自己提出的降维方法,既保存了判别信息,又能去噪声,并在实验部分与其他降维方法进行了比较。

3) SVM,做判别。

实验:Baseline (Siagian and Itti )

2.4 Kaiqi Huang (Institute of Automation)

Publication:

Biologically Inspired Features for Scene Classification in Video Surveillance(SMC 11)

主要内容:

用到Gabor,涉及场景分类比较多,也用到了显著性的概念。

3. 自己的实验

3.1 先提显著性图,再做Gabor变换。想加快速度并提高识别率。结果速度有提升,识别率有些类别提升了,部分类别下降了。

3.2 用std替换max操作。识别率在Caltech 101库上下降了。

3.3 用EBIM中的梯度阈值做了实验:

BIM(biologically inspired model)小结 - 淮静 - 淮静的博客

速度是之前的两倍,识别率并没有稳定提升。

4. 心得

自己的积淀不够,就只能做一些“偷梁换柱”的工作。不过,Guodong Guo只是将BIM用在年龄估计上,改了改参数,就发了CVPR,着实让我小小吃惊了一下。

花了一个半月时间在BIM上,本来还想加入特征的空间关系或者在pooling上做一些工作,现在基本决定放弃了。一方面:BIM这个框架有点老,现在好的框架很多。另一方面:想找一些细一点的方向,这样好突破。

BIM的意义:

a. BIM对位置和尺度变化有一定的容忍度,这个值得借鉴。

b.如果能像Guodong Guo那样找到BIM好的应用,性能的确比state-of-art的好,也是一个突破口。

你可能感兴趣的:(BIM(biologically inspired model)小结)