Gender Recognition from Unconstrained and Articulated Human Body

Qin Wu, Guodong GuoThe Scientific World JOURNAL 2014 · DOI

TL;DR

This paper tackles the challenging problem of recognizing gender from unconstrained, articulated human body images, moving beyond traditional face-based methods. They propose a novel approach that leverages body part detection, normalization, and feature extraction, demonstrating that gender can be reliably inferred even from partial or occluded body views.

What Problem Does This Paper Address?

Gender recognition is a valuable task with applications in areas like business intelligence, image search, and security. Historically, this has been approached by analyzing facial features. However, faces can be problematic for recognition due to low resolution, motion blur, or challenging viewpoints. This paper addresses the limitations of face-based gender recognition by exploring the potential of using the human body itself as the primary source of information. The core challenge is to develop a robust system that can handle the vast variability in human body poses (articulations), occlusions, and unconstrained imaging conditions, where traditional methods relying on upright, whole-body images fall short.

Methods

The authors propose a multi-stage approach for gender recognition from articulated human bodies:

  1. Upper-Body Detection and Part Localization: They first detect the upper body and then localize individual body parts (e.g., torso, head, arms) using a part-based model and pictorial structure models. This is crucial for aligning and processing body parts consistently.
  2. Body Part Normalization: Detected body parts are normalized to a standard size and orientation. This step is vital for ensuring that features extracted from different body parts across various images are comparable, regardless of their original scale or pose.
  3. Feature Extraction: Several feature representations are explored for each normalized body part, including Histograms of Oriented Gradients (HOG), Local Binary Patterns (LBP), Scale-Invariant Feature Transform (SIFT), and RGB color histograms.
  4. Data Fusion: To improve robustness and accuracy, two data fusion schemes are investigated:
  5. Feature-level fusion: Concatenating features from multiple body parts into a single, larger feature vector.
  6. Decision-level fusion: Combining the classification decisions from individual body parts or feature sets using majority voting.
  7. Partial Least Squares (PLS) Estimation: PLS is explored for both dimensionality reduction and direct classification. This method aims to efficiently reduce the high-dimensional feature space while retaining discriminative information.

Datasets

The paper utilizes two publicly available, unconstrained human body datasets:

  • Buffy: Video frames from the TV show "Buffy: The Vampire Slayer." This dataset is characterized by uncontrolled conditions, cluttered backgrounds, dark illumination, and significant scale variations.
  • PASCAL VOC 2008 challenge: This dataset consists of amateur photos with challenging illumination, low image quality, and diverse articulations.

Both datasets provide annotated stickmen, which are used in "Case I" experiments to evaluate performance without the confounding factor of body part detection errors. "Case II" experiments evaluate the system's performance with automatically detected body parts.

Key Insights and Ideas

  • Body Parts Beyond the Face: The research strongly supports the idea that human bodies, even in articulated and partial views, contain significant information for gender recognition, often surpassing the reliability of facial cues in unconstrained settings.
  • Informative Body Parts: Not all body parts contribute equally. The study reveals that lower arms are largely uninformative for gender recognition, while the torso, head, and upper arms are the most discriminative.
  • The "More is Not Always Better" Principle: Combining more body parts doesn't always lead to higher accuracy. This is attributed to the varying informativeness of parts, detection reliability, and occlusion issues. Strategic selection of informative parts is key.
  • Data Fusion Benefits: Both feature-level and decision-level fusion schemes demonstrably improve recognition accuracy and robustness compared to using single body parts or features.
  • PLS for Efficiency: The Partial Least Squares (PLS) method proves effective for significant dimensionality reduction, reducing feature dimensions to a small, fixed number (40) without substantial loss in classification accuracy, making it suitable for real-time applications.

Key Results

  • Annotated Body Parts (Case I):
    • On the Buffy dataset, the best accuracy achieved was 98.04% using feature-level fusion with head and torso.
  • On the PASCAL dataset, the best accuracy was 73.82% using decision-level fusion on head, torso, and right upper arm.
  • Automated Gender Recognition (Case II):
    • On Buffy, the highest accuracy reached 96.11% with feature-level fusion (head, torso, right upper arm), showing a small drop from the annotated case.
  • On PASCAL, the highest accuracy was 67.74% with decision-level fusion (head, torso, right upper arm), indicating PASCAL is a more challenging dataset.
  • PLS Performance: PLS effectively reduced feature dimensions to 40, with accuracies slightly lower than non-PLS methods but offering significant computational advantages.

Limitations and Future Work

  • Dataset Challenge: The PASCAL dataset proved significantly more challenging than Buffy, suggesting that future research should focus on developing more advanced learning methods to improve accuracy on such difficult datasets.
  • Lower Arm Ineffectiveness: The study confirmed that lower arms are not informative for gender recognition, implying their exclusion in future systems.
  • Future Directions: The authors suggest exploring more advanced learning techniques to tackle the challenges presented by datasets like PASCAL, aiming for higher gender recognition accuracy in unconstrained environments.

Summary updated

Abstract

Gender recognition has many useful applications, ranging from business intelligence to image search and social activity analysis. Traditional research on gender recognition focuses on face images in a constrained environment. This paper proposes a method for gender recognition in articulated human body images acquired from an unconstrained environment in the real world. A systematic study of some critical issues in body-based gender recognition, such as which body parts are informative, how many body parts are needed to combine together, and what representations are good for articulated body-based gender recognition, is also presented. This paper also pursues data fusion schemes and efficient feature dimensionality reduction based on the partial least squares estimation. Extensive experiments are performed on two unconstrained databases which have not been explored before for gender recognition.
BibTeX
@article{wu2014gender,
  author = {Qin Wu and Guodong Guo},
  title = {{Gender Recognition from Unconstrained and Articulated Human Body}},
  journal = {The Scientific World JOURNAL},
  year = {2014},
  doi = {10.1155/2014/513240},
}
Figure