Object Detection with Discriminatively Trained Part Based Models

Presentation on theme: "Object Detection with Discriminatively Trained Part Based Models"— Presentation transcript:

1 Object Detection with Discriminatively Trained Part Based Models
Pedro F. Felzenszwalb, Ross B. Girshick, David McAllester, and Deva Ramanan

2 Motivation Problem: Detecting and localizing generic objects from categories (e.g. people, cars, etc.) in static images. Issues to overcome: Changes in illumination or viewpoint Non-rigid deformations, e.g. pose Intraclass variability, e.g. types of cars

3 Previous Works Dalal & Triggs ‘05 Fischler & Elschlager ‘73
Felzenszwalb & Huttenlocher ‘00 Histogram of Oriented Gradients (HOG) Support Vector Machines (SVM) Training Sliding window detection Pictorial structures Weak appearance models Non-Discriminative training Original Image Histogram of Oriented Gradients Pictorial Structures Model of a Face

4 Object Detection with Histogram of Oriented gradients
Combine HOG and Linear SVM Detects objects using weighted HOG filters Inspect both positive and negative weighted results Human or not? Original Image Extracted Gradient Positive Weights Negative Weights

5 Deformable Part Models (DPM)
Matching Mixture Models

6 Deformable Part Models (DPM)
Represent object by several parts Model is deformable, i.e. parts can move independently of each other Parts are “punished” for being far away from their origin

7 Deformable Part Models (DPM)
Model has a root filter FO and n part models represented by (Fi,vi,di) Fi is the i-th part filter vi is the is the origin of the i-th part relative to the root di is the deformation parameter Coarse Filter High-res Part Filter Deformation models

8 Deformable Part Models (DPM)
𝑠𝑐𝑜𝑟𝑒 𝑝 𝑜 ,…, 𝑝 𝑛 = 𝑖=0 𝑛 𝐹′ 𝑖 ∙𝜙 𝐻, 𝑝 𝑖 − 𝑖=1 𝑛 𝑑 𝑖 ∙ 𝜙 𝑑 𝑑𝑥 𝑖 , 𝑑𝑦 𝑖 +𝑏 Bias Filters Feature of subwindow at location pi Deformation Parameters Displacement of part i Score of hypothesis z… Unknown… Known… 𝑠𝑐𝑜𝑟𝑒 𝑧 =𝛽∙𝜓(𝐻,𝑧) 𝛽=( 𝐹 0 ,…, 𝐹 𝑛 , 𝑑 1 ,…, 𝑑 𝑛 ,𝑏) 𝜓 𝐻,𝑧 =(𝜙 𝐻, 𝑝 0 ,…,ϕ 𝐻, 𝑝 𝑛 ,−𝜙 𝑑𝑥 1 , 𝑑𝑦 1 ,…,−𝜙 𝑑𝑥 𝑛 , 𝑑𝑦 𝑛 ,1)

9 Deformable Part Models (DPM)
Data term Spatial info 𝑠𝑐𝑜𝑟𝑒 𝑝 𝑜 ,…, 𝑝 𝑛 = 𝑖=0 𝑛 𝐹′ 𝑖 ∙𝜙 𝐻, 𝑝 𝑖 − 𝑖=1 𝑛 𝑑 𝑖 ∙ 𝜙 𝑑 𝑑𝑥 𝑖 , 𝑑𝑦 𝑖 +𝑏 Bias Feature of subwindow at location pi Deformation Parameters Displacement of part i Filters Initial condition: 𝑑 𝑖 =(0,0,1,1) Displacement Function: 𝜙 𝑑 𝑑𝑥,𝑑𝑦 =(𝑑𝑥,𝑑𝑦, 𝑑𝑥 2 , 𝑑𝑦 2 )

10 Matching 𝑠𝑐𝑜𝑟𝑒 𝑝 0 = max 𝑝 1 ,…, 𝑝 𝑛 𝑠𝑐𝑜𝑟𝑒( 𝑝 0 ,…, 𝑝 𝑛 )
The overall score of a root location is computed according to the best possible placement of parts High scoring root locations define detections High scoring part roots define object hypothesis 𝑠𝑐𝑜𝑟𝑒 𝑝 0 = max 𝑝 1 ,…, 𝑝 𝑛 𝑠𝑐𝑜𝑟𝑒( 𝑝 0 ,…, 𝑝 𝑛 )

11 Matching

12 Mixture Models Modelling for objects is done using multiple orientations Models subject to translation and rotation around the axis perpendicular to the page

13 Mixture Models Models are compared to source images in parallel
Scores of model and part filtering are combined for detection

14 Latent SVM Add your first bullet point here
Add your second bullet point here Add your third bullet point here

15 Training Add your first bullet point here
Add your second bullet point here Add your third bullet point here

16 Results (PASCAL VOC 2008) Seven total systems competed
DPM placed first in 7/20 categories

17 Title and content layout with SmartArt
Task description Step 1 Title Step 2 Title Step 3 Title

18 Title and content layout with chart

19