Discriminative part-based models. Many slides based on P. Felzenszwalb

Transcription

1 More sliding window detection: ti Discriminative part-based models Many slides based on P. Felzenszwalb

2 Challenge: Generic object detection

3 Pedestrian detection Features: Histograms of oriented gradients (HOG) Partition image into 8x8 pixel blocks and compute histogram of gradient orientations in each block Learn a pedestrian template using a linear support vector machine At test t time, convolve feature map with template t HOG feature map Template Detector response map N. Dalal and B. Triggs, Histograms of Oriented Gradients for Human Detection, CVPR 2005

4 Discriminative part-based models Root Part Deformation filter filters weights P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan, Object Detection with Discriminatively Trained Part Based Models, PAMI 32(9), 2010

5 Object hypothesis Multiscale model: the resolution of part filters is twice the resolution of the root

6 Scoring an object hypothesis The score of a hypothesis is the sum of filter scores minus the sum of deformation costs Subwindow n features n Displacements n i i i ( dxi, dyi, dxi, dyi i= 0 i= 1 score( p,..., p ) = F H ( p ) D ) Filters Deformation weights

7 Scoring an object hypothesis The score of a hypothesis is the sum of filter scores minus the sum of deformation costs Subwindow n features n Displacements n i i i ( dxi, dyi, dxi, dyi i= 0 i= 1 score( p,..., p ) = F H ( p ) D ) Filters Deformation weights Recall: pictorial structures E( l1,..., ln) = mi ( li ) + dij ( li, l j ) i Matching cost i, j Deformation cost

8 Scoring an object hypothesis The score of a hypothesis is the sum of filter scores minus the sum of deformation costs Subwindow n features n Displacements n i i i ( dxi, dyi, dxi, dyi i= 0 i= 1 score( p,..., p ) = F H ( p ) D ) Filters Deformation weights score e( ( z ) = w H ( z ) Concatenation of filter and deformation weights Concatenation of subwindow features and displacements

9 Detection Define the score of each root filter location as the score given the best part placements: score( p ) = max score ( p,..., p ) ( 0 0 p,..., p 1 n n

10 Detection Define the score of each root filter location as the score given the best part placements: score ( p0 ) = max score ( p0,..., p 1,..., p Efficient computation: generalized distance transforms For each default part location, find the bestscoring displacement n p n ) R i ( x, y) = max dx, dy ( 2 2 F H ( x + dx, y + dy) D ( dx, dy, dx, dy )) i i Head filter Head Distance filter transform responses

11 Detection

12 Matching result

13 Training Training data consists of images with labeled bounding boxes Need to learn the filters and deformation parameters

14 Training Our classifier has the form f ( x ) = max w H ( x, z ) z w are model parameters, z are latent hypotheses Latent SVM training: Initialize w and iterate: Fix w and find the best z for each training example (detection) Fix z and solve for w (standard SVM training) Issue: too many negative examples Do data mining to find hard negatives

15 Car model Component 1 Component 2

16 Car detections

17 Person model

18 Person detections

19 Cat model

20 Cat detections

21 Bottle model

22 More detections

23 Quantitative results (PASCAL 2008) 7 systems competed in the 2008 challenge Out of 20 classes, first place in 7 classes and second place in 8 classes Bicycles Person Bird Proposed approach Proposed approach Proposed approach