Frozen Frequency Disorder for Oscillator-Based Motion Segmentation of Event Streams

Event cameras do not produce frames. Each pixel independently reports a timestamped "brightness changed here" signal, so a moving scene arrives as a sparse cloud of points in space and time. Grouping those points by which object produced them, called motion segmentation, is currently done by supervised graph networks trained on labeled data. This proposal instead treats each small patch of events as a coupled oscillator: a unit that rotates at its own rate and tends to fall into step with its neighbours. Each oscillator's rate is set by the local motion it measures. Patches on the same moving object then lock into a shared rhythm, while patches on objects moving differently drift apart. Segments are read off as groups that stay in step, which is a clustering method that runs incrementally as events stream in.

The twist comes from physics. Oscillator theory says that identical units on regular, symmetric networks with communication delays often fail to lock. They settle into spatial phase waves, or they are destabilized by the delays. Adding a small amount of frozen, random mismatch between units can rescue locking. The surface of a single rigidly moving object in an event stream is exactly that bad case: near-identical local motions, a lattice-like neighbour graph, and staggered, delay-like update times caused by asynchronous event arrival. We claim that deliberately adding calibrated random frequency mismatch inside each object will make objects lock as single segments more reliably. The mismatch must be frozen, meaning fixed per oscillator, rather than fresh noise at every step. This transfers asymmetry-induced synchronization from electronic circuits to perception, and it can be falsified on a public benchmark.

Core Claim

Consider an oscillator network built over a spatio-temporal event cloud. Natural frequencies are set by local plane-fit velocity. Coupling is bounded (sinusoidal) along a neighbour graph that has been filtered to follow the local event surface. For this network we claim three things. First, segmentation quality as a function of added frozen frequency disorder is non-monotone, with a maximum at a non-zero disorder level. Second, that optimum beats an equal-variance time-varying noise injection. Third, random disorder at the optimum matches or beats disorder patterns optimized by gradient descent.

The first prediction comes from two cited sources pulling in opposite directions:

The research file flags this contradiction as unaddressed. Our resolution is that the two results govern different quantities. Mismatch helps the dynamics escape bad configurations, meaning spatial phase waves and delay resonances. The spectral bounds cap how much mismatch the locked state can absorb. The useful window is therefore bounded above by the Dörfler–Bullo condition within each object and below by the delay and symmetry pathologies. Separating two objects additionally requires that the frequency gap between them exceed what the coupling can bridge.

Mechanism

Each oscillator $i$ is a pair of phases $\theta_i \in \mathbb{T}^2$, one per image axis. It evolves as

$$\dot{\theta}_i = a\,v_i + \sigma\,\xi_i + \frac{K}{d_i}\sum_{j \in \mathcal{N}(i)} \sin(\theta_j - \theta_i),$$

where the terms are as follows:

  • $v_i$ is the 2D velocity from eigen-decomposition of the local $(x,y,t)$ event covariance, as in Powerline Tracking with Event Cameras.
  • $a$ is a gain mapping velocity to frequency.
  • $\xi_i$ is a standard normal vector drawn once, when the oscillator is created, and then frozen.
  • $\sigma$ is the disorder level under study.
  • $K$ is the coupling strength and $d_i$ the degree of node $i$.
  • $\mathcal{N}(i)$ is the node's neighbours in a surface-aware graph. That graph keeps only edges tangent to the local motion surface, following Learning Visual Motion Segmentation Using Event Surfaces.
  • The sine acts componentwise.

Oscillators are updated only when events arrive in their neighbourhood. This gives the heterogeneous, state-dependent effective delays that make the delay-rescue results relevant.

Segments are the connected groups whose time-averaged frequencies agree within a tolerance. The bounded sine coupling is what makes this more than smoothing. Linear diffusion averages across a boundary no matter how large the velocity jump. A sine coupling cannot pull together two groups whose frequency gap exceeds the coupling, so object boundaries stay sharp.

Research Program

Stage one: controlled mechanism tests. We build synthetic event streams in which the ground-truth symmetry is known exactly:

  • translating bars;
  • a rotating textured disk;
  • two overlapping translating planes forming the T-junction occlusions described in the event-surface work.

On these we measure the following as functions of $\sigma$, $K$ and graph degree:

The key diagnostic is whether identical-frequency object surfaces actually fall into spatial phase waves or delay-driven drift, and whether frozen disorder removes those failures. If identical surfaces already lock reliably, the asymmetry premise is wrong for this data, and the project reduces to an oscillator clusterer without the physics claim. We would report that outcome.

Stage two: disorder versus design. We learn the per-oscillator offsets $\xi_i$ by backpropagating through the unrolled dynamics against a segmentation loss. We then compare these learned offsets with random draws of matched variance. This directly tests whether "randomness outperforms design" survives in a learning setting where design means gradient descent.

Stage three: structural asymmetry. We test the structural version of the effect. Topological Control of Synchronization Patterns: Trading Symmetry for Stability shows that for most symmetric clusters of identical nodes the most stable topology is asymmetric; it found such topologies by simulated annealing, which the file notes is intractable at scale. We substitute a cheap local rule: random removal of a fixed fraction of edges, or making edges directed past-to-future to respect the causal direction of time. We compare against the degree-regular neighbour graph.

Stage four: EV-IMO. We run the full streaming system on EV-IMO, introduced in EV-IMO: Motion Segmentation Dataset and Learning Pipeline for Event Cameras. EV-IMO has up to three independently moving objects under aggressive camera egomotion. We first compensate camera motion with a global velocity estimate, so that the background and each object each form their own frequency group.

Relation to Prior Work

The closest work is Artificial Kuramoto Oscillatory Neurons. It uses learned Kuramoto dynamics on dense 2D feature maps for object binding, and reports about 59% AutoAttack accuracy on CIFAR10 against 0% for a ResNet. It motivates the idea that synchronization can do clustering. It differs from our proposal in three ways:

  • it assumes homogeneous update rules;
  • it never studies disorder as a stabilizer;
  • it never touches sparse asynchronous event data.

Spectral synchronization methods in vision solve a related grouping problem in a different way. Motion Segmentation via Synchronization and MultiBodySync take a batch eigen-decomposition of correspondence matrices over frames or rigid scans. MultiBodySync reports about 3.5 seconds per pass. Both constrain our design: they show that spectral grouping is a strong reference, which is why we include a spectral baseline on the identical affinity graph.

Existing event-cloud networks are feedforward. The surface-aware graph network, SECNet and SpikePoint carry no internal oscillator state. We reuse the surface-aware graph construction as our coupling substrate.

Baselines

Naive reference. Local plane-fit velocities are clustered directly by k-means over position and velocity, with no graph dynamics at all. This bounds how much the oscillator network adds beyond its own inputs.

Strongest learned method. The surface-aware graph convolutional network of Learning Visual Motion Segmentation Using Event Surfaces. It reports 85% mIoU on EV-IMO with 117k parameters and 16 ms latency per 20 ms slice. The same paper reports 69% mIoU for PointNet++. The EV-IMO pipeline itself reports IoU of 0.83 on complex sequences. For fairness:

  • we use the published split and slice length;
  • we compare our unsupervised system directly;
  • we also compare a hybrid in which only the velocity gain $a$ and coupling $K$ are tuned on the training split, with the same hyperparameter search budget given to the network's reproduction.

Beating it means matching its mIoU within two points without per-event labels, or exceeding it in the hybrid setting. Measured per-event update cost must stay below its 16 ms slice latency.

Strongest classical grouping method. Normalized spectral clustering on exactly the same surface-aware affinity graph, with velocity-difference edge weights and the number of clusters given. This matches the eigen-decomposition core of the synchronization-based segmenters. It runs on the same slices and uses the same velocity estimates.

Ablations isolating the key ingredient. The claim rests on these, all run with identical graphs and budgets:

  • $\sigma = 0$, the identical-surface network;
  • equal-variance fresh noise redrawn at every step, as in a stochastic Kuramoto model;
  • learned offsets;
  • random versus causal-directed versus degree-regular coupling graphs.

Why not just…?

Why not just cluster the local velocity estimates, or run spectral clustering on the same graph? This is the most serious threat. If spectral clustering on our affinity matches our mIoU, the oscillator dynamics contribute nothing to accuracy. The remaining argument would then be purely streaming cost: local updates at O(edges) per event, against repeated global eigen-decompositions, which are what make MultiBodySync take seconds per pass. We report both baselines head-to-head. If spectral clustering wins on accuracy, we will say so plainly and scope the contribution to streaming efficiency and the disorder finding.

Why not just smooth velocities over the graph with a bilateral or mean-shift filter? An edge-preserving smoother is the closest non-oscillatory analogue of bounded sine coupling. It shares the property that large velocity jumps do not get averaged. What it lacks is a locking transition: there is no analogue of an object snapping into one group once coupling exceeds within-object spread. So it has nothing for disorder to act on. We include it at matched neighbourhood size. If it equals the $\sigma = 0$ oscillator, the dynamics add nothing beyond smoothing. The interesting question is then whether disorder lifts the oscillator above the smoother.

Why not just add noise, which is ordinary regularization? Asymmetry-induced synchronization is specifically about quenched mismatch. The mode-mixing account in the Chua-circuit work depends on fixed off-diagonal coupling between perturbation modes, which time-varying noise does not provide. The fresh-noise ablation at equal variance settles this directly. If it performs as well as frozen disorder, the physics framing fails and we will report that.

Why not just train the graph network harder, since it already reaches 85%? Supervision is expensive on event data, where EV-IMO's labels required VICON motion capture. The supervised network also has no mechanism for trading extra computation at test time for accuracy, whereas oscillator dynamics can be run longer, as AKOrN's out-of-distribution Sudoku result (89.5% against 30.3% for a recurrent Transformer) suggests. The honest claim is competitiveness without labels, not dominance over supervised training.

Decisive Checks

The project succeeds on its central claim if all three of the following hold on EV-IMO and on the synthetic scenes, averaged over at least ten random seeds for $\xi$:

  • a sweep of $\sigma$ shows an interior optimum that improves mIoU and single-cluster lock probability over $\sigma = 0$ by more than the seed-to-seed spread;
  • frozen disorder beats equal-variance fresh noise;
  • random offsets at the optimum are not beaten by learned offsets beyond seed variance.

We further predict that the optimal $\sigma$ scales with $K$ and sits below the within-object limit $\lambda_2(L) > \|B^T\omega\|_2$. This can be checked by computing the bound per object on the synthetic scenes, where both quantities are known exactly.

The claim fails outright in either of two cases:

  • identical-surface networks already lock without spatial phase waves or delay drift;
  • mIoU decreases monotonically in $\sigma$.

In that case the heterogeneity-hurts bound governs perception, and the rescue effect seen in circuits does not transfer.