TL;DR
This paper introduces Fractual Projection Forest (FPF), a novel non-deep learning pipeline for point cloud classification. FPF leverages fractal features to achieve competitive accuracy with deep learning models while being significantly faster and more interpretable, offering a promising alternative for applications where these factors are critical.
What Problem Does This Paper Address?
Point clouds, irregular 3D data representations, are increasingly vital in fields like autonomous driving and robotics. While deep learning (DL) models have achieved high accuracies in point cloud classification, they often rely on complex, stacked modules that lead to prolonged processing times and, crucially, a lack of interpretability. This opacity is a significant hurdle in safety-critical applications where understanding why a model makes a decision is paramount. Existing explainability methods for DL models are still in their infancy and can be unreliable. The paper aims to bridge this gap by proposing a method that sacrifices minimal accuracy for substantial gains in speed and interpretability, moving away from the "arms race" of increasingly complex DL architectures.
Methods
The core of FPF lies in its novel approach to feature extraction. Instead of relying on deep neural networks, FPF converts point cloud classification into a traditional machine learning task. The process involves:
- Projection: The 3D point cloud is projected onto the three primary axes (x, y, z) and the three principal planes (xy, xz, yz), resulting in 1D and 2D representations.
- Fractal Sampling: For each projection, fractal sampling is performed using multi-size fractal windows. This technique, inspired by Hausdorff dimension, captures the geometric smoothness and distribution of points within these windows. The window sizes are varied exponentially to ensure smooth transitions.
- Feature Generation & Concatenation:
- Global Features: These include the count of non-zero fractal windows, basic statistics (min, max, index) within windows, and parameters from fitting Gaussian distributions to point distributions within windows.
- Local Features: These capture averaged distributions of points within all windows and specific monitored windows.
- These global and local features are then concatenated.
- Prediction: A traditional machine learning model, specifically a Random Forest classifier, is trained on these extracted fractal features.
- Augmentation: FPF incorporates rotation augmentation, creating multiple versions of the input data by rotating the object and extracting features from each, enhancing rotation invariance.
- Attribution Filtering: This is a unique post-processing step where features identified as having negative attributions (diminishing confidence) are filtered out, further enhancing accuracy.
Datasets
The paper evaluates FPF on the following datasets:
- ModelNet40: A standard benchmark for point cloud classification, containing 9,843 CAD models across 40 classes.
- ScanObjectNN: A real-world dataset with 15,000 objects in 15 categories, notable for containing incomplete objects and only surface information. The paper uses the "PB T50 RS" variant, considered the hardest.
- ShapeNet: Originally for 3D reconstruction, it's extended for classification tasks in this paper.
Key Insights and Ideas
- Fractal Features as a Viable Alternative: The paper demonstrates that carefully engineered fractal features can rival the performance of complex deep learning models for point cloud classification, challenging the notion that only DNNs can achieve high accuracy.
- Speed and Interpretability Trade-off: FPF offers a compelling trade-off, achieving significantly faster inference and training times (minutes on a CPU) compared to DNNs, while also providing inherent interpretability.
- Intrinsic Explainability: Unlike post-hoc methods for DNNs, FPF's structure allows for more direct explanations, such as Gini Importance and Grouped Feature Ablation, which are more reliable and easier to evaluate.
- Attribution Filtering's Impact: The post-processing step of filtering out negative attribution features significantly improves accuracy, highlighting the importance of feature selection even in a non-DL context.
- Limitations of Traditional Explainability: The paper implicitly shows that even with interpretable models, challenges remain in explaining feature importance, particularly with correlated features or when dealing with complex real-world data.
Key Results
- ModelNet40: FPF (filtered) achieves 86.3% OA, which is within 4.1% and 7.5% of PointNet and PointMLP, respectively. Crucially, FPF predicts the entire test set in under 1 second on a CPU, with training taking around 1 minute.
- ScanObjectNN: FPF outperforms PointNet on the challenging ScanObjectNN dataset, achieving 68.8% OA (filtered) compared to PointNet's 68.0%.
- Speed: FPF is orders of magnitude faster than DL models. For example, it predicts ModelNet40 test set in <1 second on a CPU, while PointNet++ takes ~1.2 seconds per instance on a GPU.
- Model Size: FPF models are significantly smaller (e.g., 76.1 MB for FPF vs. 39.8 MB for PointNet, but PointHop is 3.16 GB).
- Interpretability: The paper provides visual explanations (Figures 3 and 4) showing feature attributions, which are more intuitive than those from DNNs.
Limitations and Future Work
- Accuracy Gap: While competitive, FPF still has a slight accuracy gap compared to the state-of-the-art DNNs on ModelNet40.
- Feature Engineering: The performance of FPF can be sensitive to the handcrafted features, and optimizing these for specific datasets (like ScanObjectNN) is an area for future work.
- Rotation Augmentation: The paper notes that rotation vote techniques, while beneficial for some point cloud models, did not boost FPF's performance and even caused a collapse when rotation augmentation was not employed. This suggests that fractal features change significantly with rotations.
- Explainability Nuances: The paper acknowledges the inherent difficulties in explainability, including the lack of ground truth and potential biases in attribution methods. Future work could focus on developing more accurate and plausible explanations.
- Extensibility: The authors suggest that FPF is extensible and can incorporate more appropriate features for different datasets, opening avenues for further research.
Summary updated