Abstract
Point cloud processing architectures like PointNet are highly sensitive to outliers. This paper introduces Attentive Context Normalization (ACNe), a novel normalization layer that uses learned attention weights to exclude outliers from global statistic calculations. By mimicking robust iterative optimization, ACNe achieves state-of-the-art performance in 3D classification and camera pose estimation, maintaining high accuracy even with outlier ratios exceeding 80%.
Introduction
In computer vision, point clouds—unordered collections of coordinates—are essential for representing 3D sensor data and sparse correspondences. While architectures like PointNet have revolutionized this field, they share a significant vulnerability: extreme sensitivity to outliers. Most of these networks rely on Context Normalization (CN), which shifts and scales feature maps based on global mean and variance. Mathematically, calculating these statistics is equivalent to a least-squares problem, which is notoriously fragile. A single outlier can pull the mean away from the true data distribution, poisoning the global context. In real-world tasks like wide-baseline stereo matching, outlier ratios can exceed 80%, causing standard networks to fail.
Methods
The authors propose Attentive Context Normalization (ACN) to bake robustness directly into the network's normalization layers. Instead of standard statistics, ACN computes a weighted mean and variance, effectively allowing the model to "ignore" noise.
As shown in the architecture diagram, the process involves:
- Attentive Weighting: The network learns a weight for every point using local attention (individual point features) and global attention (relative ranking via softmax).
- Robust Normalization: These weights are used to calculate statistics, suppressing the influence of likely outliers.
- Iterative Refinement: By stacking these into Attentive Residual Blocks (ARB), the network mimics Iteratively Re-weighted Least Squares (IRLS). As data passes through deeper layers, the weights are refined, allowing the network to progressively "zoom in" on the true signal while remaining permutation-equivariant.
Results and Discussion
The authors evaluated ACNe across synthetic and real-world domains, including robust line fitting, point cloud classification (MNIST and ModelNet40), and camera pose estimation.
Key performance metrics include:
- Robust Line Fitting: At an 80% outlier ratio, ACNe achieved an error of 0.024, whereas the standard CNe error was more than double at 0.056.
- Noisy MNIST: With 60% "salt-and-pepper" noise, ACNe maintained 93.7% accuracy, while standard PointNet performance plummeted to 54.8%.
- Camera Pose Estimation: On challenging outdoor stereo benchmarks, ACNe outperformed the previous state-of-the-art (OANet) by 14.1% in mean Average Precision (mAP) without requiring expensive post-processing like RANSAC.
Conclusion
ACNe demonstrates that weighted normalization is a superior method for global context aggregation compared to traditional global max-pooling. By preserving the identity of inliers while suppressing noise, ACNe provides a "robust-by-design" framework that is remarkably efficient, adding only ~1.5% more parameters than standard models. Future work may explore applying this attentive normalization philosophy to other deep learning domains beyond point clouds.
Summary updated