Development in Object Detection. Junyuan Lin May 4th

Size: px

Start display at page:

Download "Development in Object Detection. Junyuan Lin May 4th"

Tyrone Warner
6 years ago
Views:

1 Development in Object Detection Junyuan Lin May 4th

Line of Research [1] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection, CVPR 2005. HOG Feature template [2] P. Felzenszwalb, D. McAllester, and D. Ramanan.

2 Line of Research [1] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection, CVPR HOG Feature template [2] P. Felzenszwalb, D. McAllester, and D. Ramanan. A discriminatively trained, multiscale, deformable part model, CVPR Deformable Part Model [3] C. Desai, D. Ramanan, C. Fowlkes. Discriminative Models for Multi-Class Object Layout, ICCV Multi-class Object Relationship

3 Max-margin Formulation Linear SVM HOG Feature template Latent SVM Deformable Part Model Structural SVM Y structured output Multi-class Object Relationship

4 Dalal & Triggs Detector Concatenate HOG Features in a search window into feature vector Formulate the object detection as a binary linear classifier problem Sliding window scanning over the HOG feature pyramid Output location with highest score f(x) Possible post process (Non-maxima suppression)

5 Linear SVM classifier

6 Learned weighted Filter Training sample Positive weights Negative weights *

7 Deformable Part Model Multiscale model captures features at two-resolution relative to p 0 Slide by P. Felzenszwalb

8 Score of a hypothesis relative to anchor position Latent variable Slide by P. Felzenszwalb

9 Latent SVM Training Learn Semi-convexity: Hinge loss Not convex in general! is convex for negative examples is concave for positive examples For positive examples make convex by fixing the latent variable for each positive training example.

Latent SVM Training cont d Coordinate descent optimization approach: Initialize and iterate:

Fix optimize over by quadratic programming or gradient descent Model training procedure:

10 Latent SVM Training cont d Coordinate descent optimization approach: Initialize and iterate: Fix pick best for each positive example. Fix optimize over by quadratic programming or gradient descent Model training procedure: Initialization: -Root Filter : train an initial root filter F 0 using linear SVM without latent variable. -Part Filter : Initialize six parts from root filter F 0 with the same shape, place at anchor positions with most positive energy in F 0. The size of each part filter is determined to take 80% of the area of F 0. Train the latent SVM model

11 Example Results Slide by P. Felzenszwalb

12 Quantitative Result

Multi-class Object Layout Pascal Dataset: Multiple objects of

Problems with traditional binary classification + sliding window

13 Multi-class Object Layout Pascal Dataset: Multiple objects of multi-class in a single image. Problems with traditional binary classification + sliding window approach: Intra-class: Require ad-hoc non-maxima suppression as post processing. Inter-class: multiple class models are searched independently over images. Heuristically forcing mutual exclusion. Need a principal way to model spatial statistics between objects

14 Formulation Formulate as a structured prediction problem. X: a collection of overlapping windows at various scales covering the entire image Y: the entire label vector for the set of all windows in an image. spatial relationship between prediction

15 Output score of local detectors The parameter learning and inference can be formulated as a structural SVM framework. Spatial context descriptor

16 Structural SVM Multi-class SVM Joint feature map Exponential Parameters Exponential Constraints represent Y using features extract from X,Y Learn weights so that is max for correct y

17 Structural SVM cont d n-slack Formulation: (margin rescaling) 1-Slack Formulation: Slide by T. Joachims

18 Cutting-Plane Algorithm Input: REPEAT FOR Compute ENDFOR IF _ Find most violated constraint Violated by more than ε? optimize StructSVM over ENDIF UNTIL has not changed during iteration Add constraint to working set [Jo06] [JoFinYu08]

19 Theoretical Bound Theorem: Given any S is bounded by where log, the number of constraints added to working set 2 2 4R C R C ε so that is a feasible solution. Number of constraints is independent of training examples. Linear time training algorithm.

20 Summary of Structural SVM Training: (cutting plane algorithm) Prediction: Application Specific Design of Model Loss function: Representation: joint feature map Algorithm to compute:

21 Component Formulation 0-1 loss function joint feature map Greedy forward search algorithm

22 Slide by D. Ramanan

23 Slide by D. Ramanan

24 Slide by D. Ramanan

25 Slide by D. Ramanan

Object Detection with Discriminatively Trained Part Based Models

Object Detection with Discriminatively Trained Part Based Models Pedro F. Felzenszwelb, Ross B. Girshick, David McAllester and Deva Ramanan Presented by Fabricio Santolin da Silva Kaustav Basu Some slides