Enhancing Exemplar SVMs using Part Level Transfer Regularization 1 Problem Definition: Image Retrieval 2 Problem Definition: Image Retrieval query 3 Problem Definition: Image Retrieval query Retrieved Images Retrieving same category in a similar pose Image Database Example: bicycle facing left query Retrieved Images 4 A Candidate Solution: Exemplar SVM (E-SVM) [Malisiewicz’11] [Shrivastava’11] Training a SVM with a single positive and many negative samples Linear SVMs over HoG features [Dalal &Triggs’05], [Felzenszwalb’08] Exemplar SVM 5 A Candidate Solution: Exemplar SVM (E-SVM) Training a SVM with a single positive and many negative samples Linear SVMs over HoG features [Dalal &Triggs’05], [Felzenszwalb’08] Exemplar SVM Image Database Retrieval via sliding window search on the image database 6 A Candidate Solution: Exemplar SVM (E-SVM) Training a SVM with a single positive and many negative samples Linear SVMs over HoG features [Dalal &Triggs’05], [Felzenszwalb’08] Exemplar SVM Image Database Retrieval via sliding window search on the image database Retrieved Images 7 Framework: Enhanced Exemplar SVM (EE-SVM) positive sample Train E-SVM over HoG features negative samples Previously Trained Classifiers Exemplar SVM Part-Level Transfer Enhanced E-SVM 8 Benefit: Enhanced Exemplar SVM (EE-SVM) Exemplar SVM Subwindow Retrieval Query Image Retrieved Subwindows Image Database Retrieved Subwindows Subwindow Retrieval Enhanced E-SVM 9 Overview • Transfer Learning in Computer Vision – Classification & Detection • Enhanced Exemplar SVM • Feature Augmentation vs Transfer • Results & Discussion 10 Transfer Learning in Computer Vision Learning new classes by building upon previously learned classes. • Image Classification – Adaptive SVMs, – Transfer from Multiple Models, – Adaptive Multiple Kernel Learning • Object Detection – Rigid Transfer – Flexible Transfer [Yang et al. ICDM’07] [Tommasi et al. BMVC’09] [Tommasi et al. CVPR’10] [Luo et al. ICCV’11] [Duan et al. CVPR’10] [Stark et al. ICCV’09] [Aytar and Zisserman ICCV’11] [Gao et al. ECCV’12] 13 Transfer Learning for Detection Fixed Sized Transfer • Rigid Transfer [Aytar and Zisserman ICCV’11] – Transfer between fixed sized templates – Good performance, especially for smaller number of training samples. – Hard to find visually similar detectors with same aspect ratio and size. Flexible Transfer • Flexible Transfer – – – – Transfer between different sized templates. Transferring shape features [Stark et al. ICCV’09] Deformable Transfer [Aytar and Zisserman ICCV’11] Transfer via Structured Priors [Gao et al. ECCV’12] 14 Overview • Transfer Learning in Computer Vision – Classification & Detection • Enhanced Exemplar SVM • Feature Augmentation vs Transfer • Results & Discussion 15 Framework: Enhanced Exemplar SVM (EE-SVM) Train E-SVM Query Part-Level Transfer Enhanced E-SVM Previously Trained Classifiers Exemplar SVM 16 Framework: Part-Level Transfer Regularization Exemplar ui SVM 17 Parameters: Part-Level Transfer Regularization close to construction from ui’s ui close to E-SVM 18 Framework: Matching Classifier Patches Exemplar SVM ui Previously Learned Classifiers 19 Why is it beneficial? Part-Level Transfer Regularization • Part level transfer is beneficial because… – parts can be relocated (deformation), – the possibility of finding a good match for transfer increases when we look at smaller classifier patches. • Advantages of transferring parts from well trained classifiers: – Better background suppression and discriminativity due to well trained source classifiers. – Better handling of local variations since source classifiers are trained on many positive samples. • No additional cost on runtime 20 Where is it beneficial? Part-Level Transfer Regularization • Unusual Poses • Composition of Objects [Visual Phrases - Sadeghi CVPR’11] 21 PASCAL 2007: Results - Left Facing Horse query Enhanced E-SVM E-SVM 22 PASCAL 2007: Results - Left Facing Bicycle query Enhanced E-SVM E-SVM 23 PASCAL 2007: Visual Phrase – Riding Horse query Enhanced E-SVM E-SVM 24 ImageNet: Unusual Pose - Bicycle query Enhanced E-SVM E-SVM 25 Overview • Transfer Learning in Computer Vision – Classification & Detection • Enhanced Exemplar SVM • Feature Augmentation vs Transfer • Results & Discussion 27 Implementation: Transfer vs. Feature Augmentation .... Transfer Regularization is equivalent to learning . 0.2 with augmented 0.7 0.1 features. “normal” SVM . . 29 Implications: Transfer vs. Feature Augmentation • This equivalence is not specific to Exemplar SVMs. • Transfer regularization can be implemented as feature augmentation. • Transfer regularization can be efficiently solved using standard SVM packages. 30 Overview • Transfer Learning in Computer Vision – Classification & Detection • Enhanced Exemplar SVM • Feature Augmentation vs Transfer • Results & Discussion 31 PASCAL 2007: Quantitative Results 32 ImageNet: Quantitative Results • Three queries are evaluated for each of the five classes. • Precisions at top 5, 10, 50 and 100 are reported. 33 EE-SVM E-SVM Query Handling Occlusions 34 EE-SVM E-SVM Query Handling Truncation 35 Conclusions • Boosted the performance of E-SVM which incurs no additional cost on runtime. • Presented the equivalence between Transfer regularization and feature augmentation. • Showed the benefit for unusual poses and visual phrases. • Handling truncation and occlusion. 36