Introduction IFine-grained Recognition anna’s hummingbird

ruby-throated hummingbird

Human Attribute PredictionI

Pose-normalized representations [1]I

Deformable Part Model (DPM) Weakly supervised DPMI I Fix-sized part filters initialized by heuristics. I Components initialized by clustering aspect ratio. I Strongly supervised DPM [2] I Semantic part filters initialized by part annotations. I Clusters pose information to initialize the components. Computational efficient DPM detections [3].I I Strong DPM provides semantic part localizations for pose-normalized representations. I What about simpler weak DPM without pose annotations?

Method Deformable part descriptors (DPD) Test Image Part Localization Pose-normalization Classification

The first descriptor (top row) applies a strong DPM for part localization then pool features from these inherently semantic parts.I I The second descriptor employs a weakly supervised DPM for part localization and then used a learned semantic correspondence weights to pool features from the latent parts into semantic regions. How Weights Get Computed (j) Iw il∈ W of size |P| × |R| × |C|. part of component c (j). rl : semantic region. keypoints or other semantic labels. I ρkl ∈ [0, 1]: relevance of ak to region rl . I Ijk : training images with ak and component c (j).(j) I p : i-th i I ak ∈ A:

Pooling/Classification 1Pose-normalized representationI 36Pooled image feature for semantic region Ψ(l, rl ).8 1 vs all linear SVM using Ψpn for final classification. Example Results and Failure TORSOCases Top scored people with long hair.

LEGS0.0%0.2%32.1%0.2%10.6%18.7%8.2%29.9%Top scored people wearing long sleeves.

Most confused failure case of males.

Experimental Results Fine-grained Recognition

Results on CUB200-2010 dataset .

Results on CUB200-2011 dataset. Human Attribute Prediction

Results on the Human Attributes dataset. Localization Results of strong DPM Samples of correct part localizations.

Failure cases of part localizations.

References [1] Ning Zhang, Ryan Farrell and Trevor Darrell. Pose Pooling Kernels for Sub-Category Recognition. In CVPR 2012. [2] Hossein Azizpour and Ivan Laptev. Object Detection Using Strongly-Supervised Deformable Part Models. In ECCV 2012. [3] Charles Dubout and Franc¸ois Fleuret. Exact Acceleration of Linear Object Detectors. In ECCV 2012. [4] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng and Trevor Darrell. DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition. On Arxiv.