poselets: body part detectors trained using 3d human pose annotations zuo zhen 27 sep 2011
DESCRIPTION
Introduction The proposed poselet classifiers are directly trained to handle the visual variation associated with a common underlying semantics.TRANSCRIPT
Poselets: Body Part Detectors Trained Using 3D Human Pose Annotations
ZUO ZHEN27 SEP 2011
Outline
• Introduction• Related work• Methods• Experiments• Conclusion and future work
Introduction
The proposed poselet classifiers are directly trained to handle the visual variation associated with a common underlying semantics.
Introduction• What is poselet?A poselet describes a particular part of the human pose under a givenviewpoint. It is defined with a set of examples that are close in 3D configuration space.
• Two criteria of “good” Poselets1. Easy to find the poselet given the input image. (Tightly clustered in appearance space)2. Easy to localize the 3D configuration of the person conditioned on the detection of a
poselet. (Tightly clustered in configuration space)
• Contribution1. Propose a new notion of part, a “poselet”, and an algorithm for selecting good poselets.2. Develop a novel dataset H3D(Humans in 3D) which is annotated with 3D configuration
information.
Related work1. Work in the pictorial structure traditionDisadvantage: most natural to construct kinematic simulations of a moving person, while may not correspond to the most salient features for visual recognition.
2. Work in the appearance based window classification tradition
Disadvantage: not suitable for pose extraction or localization of the anatomical body parts or joints.
3. Work of hybrid approach which have stages of one type followed by a stage of another type
Disadvantage: the parts themselves are not jointly optimized with respect to combined appearance and configuration space criteria
Left Hip
Left Shoulder
Method
This paper use keypoints to annotate the joints, eyes, nose, etc. of people to find correspondence at training time
Method(H3D dataset)
• H3D dataset: 2000 human annotations Images from Flickr with Creative Commons
Attributions License4. Provides annotation of 15 types of regions of a
person, and 19 types of keypoint annotations.
Method (H3D dataset)
• Why 3D not 2D?3D 2D
Use ratio of annotations contribute to the statistics
Every annotation Only frontal view annotations
Sensitivity to foreshortening
Not strongly affected Strongly affected
Whether allow for decomposing camera view point
Yes NA
Whether allow for query for the appearance of poselets
Yes NA
Method (H3D dataset)
Left: H3D can generate conditional region probability masks. Right: H3D can generate scatter plots of the 2D screen locations of the right elbow and left ankle given the locations of both shoulders.
Method (Finding Candidates)
Define the (asymmetric) distance in configuration space from example s to example r as:
Where = [x, y, z] are the normalized 3D coordinates of the i-th keypoint of the example s. The weight term isa Gaussian with mean at the center of the patch. The term is a penalty based on the visibility mismatch of keypoint i in the two examples.
, ( )s rh i
( )sX i( )sw i
Method (Generate Poselet Candidates)
Example query regions (left column) and the corresponding closest matches in configuration space generated by H3D.
Method (Training Poselet classifiers)
1. Given a seed patch2. Find the closest patch (search by running a scanning
window over all positions and scales of all annotations)
3. Sort them by residual error4. Threshold them5. Select a small set of poselets that are: Individually
effective and complementary6. Use them as positive training examples to train a
linear SVM with HOG features
( )sd r
Method (For Detection & Localization)
The probability of detecting the object O at position x is:
Where is the score that a poselet classifier assigns to location x and is the weight of the poselet, and the author use the Max Margin Hough Transform to learn the weight.
( )ia x
iw
Experiments
(1) Detecting Human Torsos
ROC curve comparing the proposed torso detection performance together with other published detectors on the H3D test set
Experiments
• Examples of torso detections using poselets
Experiments
(2) Detecting People on PASCAL VOC 2007Outperform the part-based deformable detector on H3D but get comparable performance on VOC2007.
Experiments
(3) Detecting Keypoints
Detection rate of some keypoints conditioned on true positive torso detection.
Conclusion & Future Work
• ConclusionThe authors propose a two-layer classification/ regression model for detecting people and localizing body components. And the 3D annotation guides the search for good parts.• Future workUse H3D more widely.
Birdlets: Subordinate Categorization Using Volumetric
Primitivesand Pose-Normalized Appearance
Outline
• Introduction• Related work• Methods• Experiments• Conclusion
Introduction
• Application backgroundCurrent research: two extremes of individuals and basic-level categoriesFew research on subordinate categorization
• What is subordinate categorization?Distinguish by the differing properties of parts.
Introduction
Overview of the Proposed approach
Introduction
• Contribution1. A framework for detecting volumetric part
models2. A pose-normalized appearance model for
comparing part appearance3. A classification model for aggregating
information about part properties
Related work
• Image featuresDisadvantages: view-dependent, pose variation• Part modelDisadvantages: high intra-class variability, significant articulation• Hierarchy modelDisadvantages: subordinate categories have both subtle and drastic appearance variation• Attribute modelDisadvantages: Insufficient to model subtle differences between parts
Method
• Why birds?1. Exist largest subordinate-level dataset (CUB-
200)2. Conform with the definition of subordinate-
level (share common structure & parts with many subtle part distinctions)
3. Involving highly variable appearances and articulations (challenging)
Method (PNAD)
Post-normalized appearance descriptor (PNAD)1. Map points on a unit sphere onto the ellipsoid’s
surface for patch sampling2. Project patches on ellipsoid surface to original
image plane3. Extend the projected patches for extracting SIFT
descriptor 4. Concatenate the location and appearance
information for forming PAND descriptor
Method (PNAD)
Method (Birdlet)
• Volumetric primitive templates1. Two parts (head & body)2. Two ellipsoids (parameters: location center,
3D orientation, scale)3. Alignment (assisted by visible point features:
beaktips, eyes, wingtips, feet and tails)
Method (Training & Testing)
1. Get selection windows for detecting objects and parts in testing image(both positive and negative examples for SVM classifier)
2. Get birdlets for integrated classification
Method (Integrated Classification)• Stacked Evidence Trees model
The Stacked Evidence Tree takes a test feature and finding a set of training features that are similar both in appearance and surface location, and ultimately returning the class label distribution across this similar set
Experiments
• Classification Confusion Matrices
(a) the PHOW/SVM Baseline (37.12% MAP), (b) the PNAD-RF performance on the top 20% of detections (40.25% MAP), and (c) the PNAD-RF performance on the ground truth part locations (66.58% MAP).
Experiments
• Example Volumetric Primitive Detections
Top two images: the bird is detected and localized with reasonable accuracyLow two images: false positive detections
Experiments
Classification of Volumetric Detections. For the k top ranked detections, this plots the corresponding PNAD-RF classification performance (using mean-average precision)
Conclusion
• ConclusionThis paper presented an approach for subordinate categorization using a pose-normalized appearance representation founded upon a volumetric part model.