Introduction
Problem: Classifying attributes and actions in still images
Model: Collection of part templates Specific scale space locations (human centric) Discriminative learning Sparse Activation
Motivation
Train Test Train Test
Overview
Image Scoring
Mining Parts &
Learning Templates
Formulation
fractional multiples of width and height
Dataset:
Model:
Objective:
Model
fractional multiples of width and height
. . . Part 1 Part 2 Part 3
parts
d = 1000 Model
Model & Scoring
Image ScoringModel
overlap constraintsparse activationOptimization: Greedy selection of 0.33 overlap constraint
Model Initialization
1) randomly sample the positive training images for patch positions:
2) Initialize model parts:
perfect case: worst case:
3) BoF features normalized 105 patches.
3) Prunning: remove unused parts
Learning
k = 4
Experiments
Willow 7 Human actions
27 Human Attributes (HAT)
Stanford 40 Human Actions
Implementation
Features:– VLFeat - Dense SIFT,
• step size: 4 pixels• square patches (8 to 40 pixels)
– k-means - vocabulary 1000– explicit feature map + Bhattacharyya (Hellinger – Square root) kernel
Baseline: 4 level spatial pyramid
Immediate context:– expand the human bounding boxes by 50% in both width and
height
Full image context:– full image classifier uses 4 level SPM with an exponential 2
kernel
Qualitative Results
Willow Actions
Database of Human Attributes (HAT)
Stanford 40 Actions
Learned Parts - I
In each row, the first image is the patch used to initializethe part and the remaining images are its top scoring patches
Learned Parts - II
In each row, the first image is the patch used to initializethe part and the remaining images are its top scoring patches
Learned Parts - III
In each row, the first image is the patch used to initializethe part and the remaining images are its top scoring patches