Strike (with) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects

Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, Anh Nguyen

Introduction

Deep Neural Networks (DNNs) have achieved remarkable success on standard benchmarks like ImageNet, often reaching superhuman performance. However, these benchmarks typically feature "stationary" distributions where objects are centered and presented in standard, canonical poses. In real-world scenarios—such as autonomous driving or robotics—AI systems frequently encounter objects in unusual orientations, such as a rolled-over vehicle or a steep aerial view. This research investigates the critical safety gap posed by these "Out-of-Distribution" (OoD) natural orientations, revealing that state-of-the-art models are surprisingly fragile when faced with strange poses of familiar objects.

Methodology

The researchers developed a framework to systematically identify "blind spots" in DNNs by integrating a 3D renderer with a target image classifier. Instead of evaluating static images, the framework treats the 3D object’s pose (rotation and translation) as parameters to be optimized to induce misclassification.

The framework connects a 3D renderer to a target DNN, using backpropagation or search strategies to find adversarial poses.

Three primary search strategies were evaluated:

  1. Random Search (RS): Sampling thousands of random orientations.
  2. Finite-Difference Gradient Descent (FD-G): Estimating gradients by testing slight perturbations in 3D space.
  3. Differentiable Rendering (DR-G): Utilizing a specialized renderer to directly optimize 3D vertices based on pixel-level influence.

Surprisingly, the simpler FD-G method proved more effective than DR-G, as the latter often produced noisy gradients that hindered optimization.

Key Insights and Results

The study utilized a dataset of 30 high-quality 3D models corresponding to ImageNet classes and tested them against architectures like Inception-v3, AlexNet, and ResNet-50. The findings highlight a profound lack of geometric understanding in current AI models.

The Fragility of Pose Space

For objects that a DNN recognizes with near-perfect accuracy in a standard view, it fails in approximately 97% of the total pose space. The "correct" classification zone is a tiny island in a sea of errors. Models demonstrated extreme sensitivity; a rotation of just 8 to 10 degrees—a minor shift in 3D space—could flip a prediction from a "school bus" to a "garbage truck" or even a "punching bag."

Examples of adversarial poses where familiar objects like school buses and fire trucks are misclassified with high confidence.

Universal Adversarial Poses

One of the most concerning discoveries is the high transferability of these failure cases. A pose that fools Inception-v3 has a 99.9% chance of also fooling AlexNet. Furthermore, 75.5% of these poses caused the YOLOv3 object detector to fail, suggesting that the flaw is rooted in fundamental feature learning rather than specific architectural quirks.

Real-World Validation

To ensure these failures weren't merely artifacts of 3D rendering, the authors collected 150 real-world photos of objects in unusual poses. 98.3% of these real-world images were misclassified by the models, confirming that the "reality gap" does not protect AI from these orientation-based errors.

Conclusion

The researchers conclude that current DNNs are "visual masochists," relying on brittle 2D pattern matching rather than a robust 3D understanding of the world. Standard adversarial training—adding these specific poses to the training set—failed to solve the problem, as the models did not generalize to new 3D models of the same class. Future progress in AI safety and robustness likely requires moving beyond larger datasets toward architectures that incorporate 3D geometric priors and perform true visual reasoning.

Summary updated

Abstract

Despite excellent performance on stationary test sets, deep neural networks (DNNs) can fail to generalize to out-of-distribution (OoD) inputs, including natural, non-adversarial ones, which are common in real-world settings. In this paper, we present a framework for discovering DNN failures that harnesses 3D renderers and 3D models. That is, we estimate the parameters of a 3D renderer that cause a target DNN to misbehave in response to the rendered image. Using our framework and a self-assembled dataset of 3D objects, we investigate the vulnerability of DNNs to OoD poses of well-known objects in ImageNet. For objects that are readily recognized by DNNs in their canonical poses, DNNs incorrectly classify 97% of their pose space. In addition, DNNs are highly sensitive to slight pose perturbations. Importantly, adversarial poses transfer across models and datasets. We find that 99.9% and 99.4% of the poses misclassified by Inception-v3 also transfer to the AlexNet and ResNet-50 image classifiers trained on the same ImageNet dataset, respectively, and 75.5% transfer to the YOLOv3 object detector trained on MS COCO.
BibTeX
@inproceedings{alcorn2019strike,
  author = {Michael A. Alcorn and Qi Li and Zhitao Gong and Chengfei Wang and Long Mai and Wei-Shinn Ku and Anh Nguyen},
  title = {{Strike (with) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects}},
  booktitle = {CVPR},
  year = {2019},
  url = {https://openreview.net/forum?id=r7-6fXxdaS},
}
Figure