TL;DR
This paper introduces a metric learning framework using Siamese Convolutional Neural Networks (CNNs) to automatically learn geometric invariants of planar curves under Euclidean and similarity transformations. By replacing traditional, hand-crafted numerical differentiation with data-driven optimization, the learned representations achieve unprecedented robustness against noise, poor sampling, and decimation compared to classical axiomatic methods.
What Problem Does This Paper Address?
In computer vision and differential geometry, analyzing planar shapes requires finding local descriptors of a curve that remain unchanged regardless of rotation, translation, or reflection. A classic example is Euclidean curvature.
However, computing these invariants from real-world, discrete data is notoriously difficult. The mathematical formulas rely on non-linear functions of high-order derivatives. When dealing with noisy sensor data or sparse sampling, numerical differentiation acts as a high-pass filter, drastically amplifying noise and rendering the calculated invariants unstable. While traditional geometry has attempted to mitigate this using "integral invariants" (low-pass filtering) or "scale-space methods" (evolving the curve over time), these remain handcrafted compromises. This paper addresses the challenge of automatically learning robust, stable geometric invariants directly from data.
Methods
The authors bridge the gap between numerical differential geometry and deep learning by utilizing a Siamese CNN configuration.
- Representation: Planar curves are explicitly represented as an $N \times 2$ array containing the $X$ and $Y$ coordinates of $N$ sampled points.
- Architecture: The network uses a multi-stage 1D CNN pipeline. It consists of three stages, each containing two sequential batches of 1D convolutions (15 filters of width 5) followed by Rectified Linear Units (ReLU) and a max-pooling unit per point. A final linear layer combines the responses into a point-wise signature.
- Loss Function: The network is trained using a contrastive loss function. Positive pairs ($\lambda = 1$) consist of a curve and its randomly rotated, translated, and reflected counterpart. Negative pairs ($\lambda = 0$) are entirely unrelated curves.
- Multi-Scale Learning: To learn hierarchical, multi-scale representations (similar to classical scale-space methods), the authors introduced a clever twist: they designed negative training pairs consisting of a curve matched with its own smoothed/evolved version (generated via local polynomial regression). This forces the network to distinguish between fine-grained local details and global geometric structures.
Datasets
The framework was trained and evaluated using the MPEG7 Database, a standard benchmark containing 1,400 shapes across 70 different categories.
- Split: 700 shapes were used for training, 350 for testing, and 350 for validation.
- Augmentation: The authors generated a training dataset of 10,000 pairs. Positive examples were created by applying random 2D rotations (sampled uniformly from $[-\pi, \pi]$), translations, and reflections. All curves were sub-sampled to 500 points and normalized by removing the mean and dividing by the standard deviation to ensure scaling invariance.
Key Insights and Ideas
- Outsourcing Filter Design: Instead of manually designing high-pass filters for differentiation or low-pass filters for integration, the optimization process naturally forces the CNN to discover filters that perfectly balance geometric fidelity with noise robustness.
- Data-Driven Scale Space: By using smoothed versions of the same curve as negative examples, the network mimics the mathematical profile of mean curvature flow (where a curve smoothly shrinks to a circle). The network learns a multi-scale representation purely from data constraints without explicitly solving differential equations.
- Geometric Interpretation of CNN Filters: The first-layer filters learned by the network resemble standard 1D Gaussian derivative kernels, providing a satisfying theoretical validation that the network is truly learning to perform robust numerical differentiation.
Key Results
The learned "Network Invariant" closely tracks the qualitative profile of axiomatic Euclidean curvature but eliminates the erratic spikes caused by noise.
- Noise Robustness: When zero-mean Gaussian noise was added to the curves, classical differential invariants completely degraded. The network invariant remained stable and highly descriptive, outperforming both differential and integral invariants at high noise levels ($\sigma = 0.3$) in a shape retrieval task.
- Resiliency to Decimation: The authors tested the signatures by downsampling a high-resolution curve's points from 70% down to just 5%. The network invariant successfully maintained its signature shape at fixed points, demonstrating remarkable resilience to sparse and irregular sampling where traditional methods failed.
Limitations and Future Work
While the paper successfully demonstrates the framework for 2D planar curves under Euclidean and similarity transformations, a few limitations and areas for future work remain:
- Extension to Higher Dimensions: The current work is strictly 1D-convolutional, operating on sequential boundary points of planar curves. Extending this to 3D surfaces or manifolds would require more complex architectures (like geometric deep learning or graph CNNs).
- Broader Transformation Groups: The model focuses on Euclidean and similarity groups. Future work could test its efficacy on more challenging transformation groups, such as affine or projective transformations.
- Theoretical Guarantees: As with many deep learning approaches, while the empirical stability is superior, the network lacks the strict mathematical uniqueness guarantees provided by Cartan's theorem for axiomatic differential invariants.
Summary updated