Presentation is loading. Please wait.

Presentation is loading. Please wait.

O BJECT D ETECTION WITH D ISCRIMINATIVELY T RAINED P ART B ASED M ODELS PRESENTED BY Xiaolong Wang.

Similar presentations


Presentation on theme: "O BJECT D ETECTION WITH D ISCRIMINATIVELY T RAINED P ART B ASED M ODELS PRESENTED BY Xiaolong Wang."— Presentation transcript:

1 O BJECT D ETECTION WITH D ISCRIMINATIVELY T RAINED P ART B ASED M ODELS PRESENTED BY Xiaolong Wang

2 D ETECTION

3 C HALLENGE Deformation Part of the Slides From Ross Girshick

4 C HALLENGE Viewpoint

5 C HALLENGE Variable structure

6 C HALLENGE Images from Chaitanya Desai

7 2-layer Model Deformable D EFORMABLE P ART M ODELS Leo Zhu, CVPR 2010

8 HOG P YRAMID Root Filter Part Filters

9 F ORMULATION One root (i=0) + n parts. Model Parameters for HOG HOG Features Model Parameters for Deformation

10 I NFERENCE

11 M ULTI - VIEWS

12 L ATENT O RIENTATION No orientation in PAMI paper (DPM v3) Use latent orientation (DPM v4) Guess what is it? right-facing horse

13 U NSUPERVISED ORIENTATION CLUSTERING

14 L ATENT O RIENTATION Inference: Choose the best view and best orientation. Learning: Train the parameters for 3 views, and flip the weights to get 3*2 views.

15 H OW IMPORTANT IT IS One view:42.1% 3-view: 47.3% 3*2-view: 56.8% For horse:

16 H OW IMPORTANT IT IS For all classes (DPM v4):

17 L EARNING Linear Formulation  Putting all features in one vector  Latent variable z represents part locations (and component index for multi-views)

18 L ATENT SVM

19 Detection on Positive Samples  Sliding window  Overlap with root-node window > 0.7

20 L ATENT SVM Hard Negative Mining Carl Vondrick HOGgles, ICCV 2013

21 L ATENT SVM Hard Negative Mining  Small or no overlap  High detection score Maintaining Sample Cache  Select no more than 500 negative samples per image;  Cache size = 20000

22 L ATENT SVM Dual Method  Not scalable. Stochastic gradient descent(DPM v4)  Important: Shuffle everytime! LBFGS(DPM v5)  Second-order Newton Method  Faster & better performance

23 3- STEP I NITIALIZATION Step-1: Only Train Root Filter  positive data (highest overlap)  No hard negative mining Car

24 3- STEP I NITIALIZATION Step-2: Merg Components  Setting root selection as latent variable

25 3- STEP I NITIALIZATION Step-3: Initialize Part Filters  Fix part number as 8 (DPM v4/5)  Sliding window, calculate L1/L2 norm of the positive weights.

26 P OST P ROCESSING Bounding Box Regression  Linear regression for (x1,y1,x2,y2) Non-Maximum Suppression  Pick up high score boxes Context

27 C ONTEXT Marr Prize 2009 Context SVM,CVPR2010 segDPM,CVPR2013

28 N UMBERS VOC 2010: 29.6 and 32.2 VOC 2007: 33.7 and 35.4 VOC 2010: segDPM(with tons of things) 40.4

29 L ARGE - SCALE D ATASET ImageNet 2013 DPM v4 in cpp

30 S UMMARY Although DPMs is loosing to CNNs, the techniques and small tricks we learned from DPMs help solving many other vision problems.

31 Q UESTIONS


Download ppt "O BJECT D ETECTION WITH D ISCRIMINATIVELY T RAINED P ART B ASED M ODELS PRESENTED BY Xiaolong Wang."

Similar presentations


Ads by Google