TimePoint: Accelerated Time Series Alignment via Self-Supervised Keypoint and Descriptor Learning

TL;DR

TimePoint tackles the bottleneck of aligning long time series by learning self-supervised keypoints and descriptors that enable DTW to operate on a sparse, informative representation. It leverages synthetic data through SynthAlign and CPAB-based nonlinear time warps to generate ground-truth correspondences, training a fully convolutional 1D architecture with WTConv layers. At test time, DTW is applied to descriptor sequences at keypoints, reducing complexity from $O(L^2)$ to $O(\tilde L\cdot\tilde L')$ and achieving substantial speedups with often improved accuracy. The approach generalizes well to real data, including ECG-like signals, and can be further enhanced by fine-tuning on real datasets. The method offers a practical, scalable solution for time-series analysis with strong robustness to noise and distortions.

Abstract

Fast and scalable alignment of time series is a fundamental challenge in many domains. The standard solution, Dynamic Time Warping (DTW), struggles with poor scalability and sensitivity to noise. We introduce TimePoint, a self-supervised method that dramatically accelerates DTW-based alignment while typically improving alignment accuracy by learning keypoints and descriptors from synthetic data. Inspired by 2D keypoint detection but carefully adapted to the unique challenges of 1D signals, TimePoint leverages efficient 1D diffeomorphisms, which effectively model nonlinear time warping, to generate realistic training data. This approach, along with fully convolutional and wavelet convolutional architectures, enables the extraction of informative keypoints and descriptors. Applying DTW to these sparse representations yield major speedups and typically higher alignment accuracy than standard DTW applied to the full signals. TimePoint demonstrates strong generalization to real-world time series when trained solely on synthetic data, and further improves with fine-tuning on real data. Extensive experiments demonstrate that TimePoint consistently achieves faster and more accurate alignments than standard DTW, making it a scalable solution for time-series analysis. Our code is available at https://github.com/BGU-CS-VIL/TimePoint

Figures (15)

• Figure 1: TimePoint (TP): Keypoint Detection and Descriptors on real-world, unseen, ECG data of length 2500 (TP was trained on synthetic data of length 512). Each panel depicts (top-to-bottom) the original signal and predicted keypoints, keypoint probability map, and PCA of the learned descriptors ($D=256,$ using 5 principal components for visualization purposes).
• Figure 2: Comparison of DTW alignment using the raw sequence (top) or TimePoint keypoints and descriptors (Bottom). The black and blue boxes highlight areas where sparse DTW using TP descriptors results in better matching.
• Figure 3: Training and Inference overview. Left: signals and keypoints are synthetically generated and augmented using CPAB warps (\ref{['Sec:Synthetic:Dataset']}). Middle: TimePoint predicts KP location and descriptors using the known correspondence (\ref{['Sec:Method']}). Right: real-world, unseen data pairs are matched using DTW on TimePoint descriptors at keypoint locations.
• Figure 4: Samples from the synthetic dataset SynthAlign.
• Figure 5: Generating signals and keypoints pairs with known correspondences using a CPAB transformation $T^\btheta$, which was obtained from a CPA velocity field, $\bv^\btheta$, as proposed in Freifeld:PAMI:2017:CPAB and briefly explained in our Appendix.