On the Robustness of Human Pose Estimation

Naman Jain, Sahil Shah, Abhishek Kumar, Arjun JainWorkshops @ CVPR 2019

TL;DR

This paper provides a comprehensive study of adversarial attacks on human pose estimation (HPE) systems, revealing that while HPE models are generally more robust than classification models, they are still vulnerable. The authors introduce new attack strategies and analyze how different architectural choices and training methodologies impact robustness, offering insights for developing more resilient HPE systems.

What Problem Does This Paper Address?

Deep learning systems are increasingly deployed in real-world applications, from autonomous driving to human-computer interaction. A critical aspect of these systems is their robustness, not only to natural variations but also to adversarial attacks – imperceptible modifications to input data designed to fool the model. While adversarial attacks have been extensively studied for image classification, their impact on human pose estimation (HPE) systems, which are crucial for applications like augmented reality and robotics, remains less understood. This paper aims to bridge this gap by conducting a thorough investigation into the robustness of various HPE architectures against different types of adversarial attacks.

Methods

The authors propose and evaluate several adversarial attack strategies tailored for HPE. These include:

  • Targeted vs. Un-targeted Attacks: Differentiating between attacks that aim to push the model towards a specific incorrect pose versus those that simply aim to increase the model's error.
  • Single-Step vs. Iterative Attacks: Comparing the effectiveness of one-shot attack methods (like FGSM) with iterative approaches (like IGSM) that apply multiple small perturbations.
  • Image-Specific vs. Universal Perturbations: Investigating attacks that are designed for a particular image versus those that are learned from a dataset and can be applied to any image.
  • White-box vs. Black-box Attacks: Examining attacks that require full access to the model's architecture and parameters versus those that only require input-output pairs.

They also analyze the impact of various design choices within HPE systems, such as:

  • Heatmap-based vs. Direct Regression: Comparing models that predict heatmaps of joint locations against those that directly regress pixel coordinates.
  • Architectural Choices: Evaluating different backbone architectures like Stacked Hourglasses (SHG) and their variations.
  • Anthropometric Constraints: Assessing how incorporating human body structure knowledge affects robustness.

Datasets

The study utilizes two widely-used datasets for human pose estimation:

  • MPII (Human Pose Estimation Dataset): Used for training and evaluating single-person 2D HPE systems.
  • COCO (Common Objects in Context): Used for evaluating multi-person 2D HPE systems.
  • Human3.6M: Used for evaluating single-person 3D HPE systems.

Key Insights and Ideas

  • HPE is More Robust Than Classification, But Still Vulnerable: While single-step attacks are surprisingly ineffective against HPE, iterative and universal attacks can significantly degrade performance, often rendering models useless.
  • Heatmap-Based Models are More Robust: Architectures that predict heatmaps tend to be more resilient to adversarial attacks compared to those that directly regress joint coordinates. This is attributed to the implicit bounding of heatmap predictions within valid image coordinates and the nature of the Gaussian bump loss.
  • Anthropometric Constraints Improve Robustness: Models that explicitly incorporate human body structure and relationships (e.g., bone-length ratios) show improved resilience to attacks.
  • Targeted Attacks are Harder: Crafting attacks to achieve a specific incorrect pose is more challenging than simply increasing the model's error, requiring more iterations and careful tuning.
  • Universal Perturbations are Highly Effective and Practical: Image-agnostic universal perturbations are shown to be very effective and computationally efficient, making them a practical threat in real-world scenarios.
  • Body-Joint Vulnerability Varies: Certain body joints, like hips and ankles, are more susceptible to attacks than others, such as the neck and head, likely due to how they are represented and localized in training data.
  • Multi-Person HPE is More Vulnerable: Multi-person HPE systems, especially bottom-up approaches, are more susceptible to attacks, leading to a significant increase in predicted human counts and a drop in performance metrics.

Key Results

  • Relative PCKh Degradation: Universal attacks can degrade relative PCKh to as low as 6.4% (training) and 9.9% (validation) for some models.
  • Iterative Attacks are More Potent: IGSM-U-100 can drive relative PCKh to less than 5%, and IGSM-T-100 can achieve targeted PCKh of ~90% even at moderate epsilon values.
  • Body-Joint Vulnerability: Hips, knees, and ankles show relative PCKh values below 5% under un-targeted attacks, indicating high vulnerability.
  • Multi-Person Performance Drop: For multi-person systems, IGSM attacks can drive MAP to near 0, with bottom-up approaches predicting significantly more humans under attack compared to top-down methods.
  • 3D HPE Vulnerability: Attacking the depth-regressor branch of 3D HPE systems can increase MPJPE by six-fold, from 60 to 360.
  • User Study Findings: While higher epsilon values lead to more perceptible differences, even non-"annoying" changes can fool the network.

Limitations and Future Work

  • Limitations:
    • The study focuses primarily on single-person 2D HPE, with limited analysis of multi-person and 3D systems.
  • Attacking the object detection module in top-down multi-person systems was done to illustrate effects, leaving end-to-end attacks for future work.
  • The study acknowledges that some results might be slightly inferior to original reported results due to using a standard protocol.
  • Future Work:
    • Explore more specialized HPE-specific attacks, such as "t-shirt attacks" or "facial feature based attacks."
    • Conduct more in-depth analyses of multi-person and 3D HPE systems.
  • Investigate design choices for improving robustness in these more complex scenarios.
  • Develop more effective defense mechanisms against adversarial attacks on HPE.

Summary updated

BibTeX
@inproceedings{jain2019robustness,
  author = {Naman Jain and Sahil Shah and Abhishek Kumar and Arjun Jain},
  title = {{On the Robustness of Human Pose Estimation}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
}
Figure