Presentation on theme: "Silhouette-based Object Phenotype Recognition using 3D Shape Priors Yu Chen 1 Tae-Kyun Kim 2 Roberto Cipolla 1 University of Cambridge, Cambridge, UK 1."— Presentation transcript:
Silhouette-based Object Phenotype Recognition using 3D Shape Priors Yu Chen 1 Tae-Kyun Kim 2 Roberto Cipolla 1 University of Cambridge, Cambridge, UK 1 Imperial College, London, UK 2
Problem Description Task: To identify the phenotype class of deformable objects. Given a gallery of canonical-posed silhouettes in different phenotype classes. Can we find out ?
Problem Description Motivation: –Pose recognition is widely investigated; –Phenotype recognition is somehow overlooked; –Applications? Difficulty: –Pose and camera viewpoint variations are more dominant than the phenotype variation.
Problem Description 2D approaches hardly work in this case. Our strategy: make use of the 3D shape prior of deformable objects. Shall we use a purely generative approach? No! Too expensive to perform for a recognition task!
Solution: Two-Stage Model Main Ideas: Discriminative + Generative Two stages: 1. Hypothesising –Discriminative; –Using random forests; 2. Shape Synthesis and Verification –Generative; –Synthesising 3D shapes using shape priors; –Silhouette verification. Recognition by a model selection process.
Use 3 RFs to quickly hypothesize phenotype, pose, and camera parameters. Learned on synthetic silhouettes generated by the shape priors. Parameter Hypothesizing F A : Pose classifier F C : Camera pose classifier F S : Phenotype classifier (canonical pose)
Examples of Tree Classifiers The phenotype classifier The pose classifier
Training RF Classifiers Random Features: –Rectangle pairs with random sizes and locations. –Difference of mean intensity values [Shotton et al. 09] –Feature error compensation for phenotype classifier; Criteria Function: –Similarity-aware diversity index.
Shape Synthesis and Verification Generate 3D shapes V –From candidate parameters given by RFs. –Use GPLVM shape priors [Chen et al.’10]. Compare the projection of V with the query silhouette S q. –Oriented Chamfer matching (OCM). [Stenger et al’03]
Comparative Approaches: Learn a single RF phenotype classifier; Histogram of Shape Context (HoSC) –[Agarwal and Triggs, 2006] Inner-Distance Shape Context (IDSC) –[Ling and Jacob, 2007] 2D Oriented Chamfer matching (OCM) –[Stenger et al. 2006] Mixture of Experts for the shape reconstruction –[Sigal et al. 2007]. –Modified into a recognition algorithm
Comparative Approaches: Internal comparisons: –Proposed method with both feature error modelling and similarity-aware criteria function (G+D); –Proposed method w.o feature error modelling (G+D–E); –Proposed method w.o similarity-aware criteria function (G+D–S) Using standard diversity index instead.
Recognition Performance Cross-validation by splitting the dataset instances. 5 phenotype categories for every test. Selecting one instance from each category.
Recognition Performance How the parameters of RFs affect the performance? –Max Tree Depth d max –Tree Number N T
Qualitative Results of SVR Left: Input image/silhouette; Centre: Using RF-hypothesizes; Right: Using the optimisation-based approach.
Qualitative Results of SVR
Take-Home Messages Phenotype recognition is difficult but still possible; Combing discriminative and generative cues can greatly speed up the inference; A divide-and-conquer strategy can help improve the recognition rate.
Future Work Explore the application on more complicated poses and more categories. –E.g. Boxing, gardening, other sports, etc. Data collection; Automate the silhouette extraction. –E.g. Kinect.
The End Questions?
Feature Error Modelling Purpose: –For learning the phenotype classifier Fs; –To reduce systematic errors between synthetic and real silhouettes. Error modelling dataset: –Several pairs of synthetic and real silhouettes with different error modes; –Compute the feature difference for each pair. Error compensation of the training data. –Find nearest neighbour silhouettes in the error modelling dataset. –Compensate the features of synthetic silhouettes with corresponding error vectors. Synthetic Silhouette Real Silhouette Error Vector e
Similarity-aware Criteria Functions Tree learning: maximize the criteria function (impurity) drop of each node Observation: some classes can be more similar to each other while some are more different. –Generalise the Gibbs and Martin’s diversity index; –Take the class similarity into account; –The similarity matrix W is defined by the average 3D mesh difference between phenotype classes. More similar Less similar
Comparative Approaches: Learn a single RF phenotype classifier. Training Data: –Generated by the 3D shape priors; –Phenotype classes are uniformly sampled from the latent space; –Various poses/camera viewpoints for each class. Recognition: –Compare the RF histograms between the query and each gallery image. –χ 2 distance. Phenotype 1 Phenotype 2 Phenotype N Random Forest...
Approximate Single View Reconstruction Use 3D shapes hypotheses V. Contrast with the results by the optimisation-based approach [Chen et al. 10]. Performance –Fairly good accuracy; –More than 50x faster than the optimisation-based approach.