The strongest gait recognizer in our file, GaitSet, reaches 95.0% rank-1 accuracy on normal walking but falls to 70.4% when subjects wear coats. GaitSet treats a walking video as an unordered bag of silhouette frames and throws away timing. When a coat hides the body's outline, static shape is exactly the information that goes missing. What a coat does not hide is the rhythm of the lower legs: how fast each shin swings and how the harmonics of the stride line up in time. We propose to capture that rhythm with the Wigner function, a physicist's tool that describes a signal jointly in time and frequency, the analog of position and momentum. We add it as a second branch beside set pooling. Because the Wigner function becomes a fixed-length, shift-invariant vector once summarized properly, the same descriptor also lets gait use the simple, training-free template-protection schemes that so far have been validated only on faces.
The sharper question underneath is one physics has already posed. A smoothed, always-positive phase-space picture (the Husimi function, which for audio is simply the Gaussian-window spectrogram) is convenient for machine learning. A sharp picture must take negative values. We ask whether the sharp, signed representation carries identity information about walking that the smoothed one blurs away. The comparison is clean and cheap, and either answer is informative.
Core Claim
Our claim is that per-body-strip motion signals from gait silhouettes, mapped into a signed time-frequency distribution and summarized by the magnitude of its Fourier transform, recover part of the CASIA-B coat-condition accuracy that set pooling loses. That summary is called the ambiguity function; it is invariant to where in the video the stride starts. We further claim that this gain comes specifically from sharpness. A spectrogram (Husimi-type) branch with the same capacity, fed the same signals, should gain less. Finally, the resulting fixed-length vector should survive data-independent template protection with hyperparameters tuned on faces, with a relative error increase comparable to what that scheme shows on faces.
Why the physics is relevant, and where it is not
Two results from the quantum side shape the design.
Symmetries and Wigner representations of operational theories generalizes Wigner functions to arbitrary measurement theories. It proves that a phase-space representation can stay non-negative only if its reference observables are jointly measurable. Time and frequency are not jointly measurable, so any representation that resolves both sharply must be signed. The positive alternative has to blur.
Q-Flow takes the opposite engineering route. It models open quantum dynamics with normalizing flows on the Husimi function precisely because positivity makes KL-based training possible, and it accepts the loss of sharpness.
Together these two papers define our central ablation. The spectrogram is the Husimi choice, and the Wigner/ambiguity branch is the signed one.
On discrete Wigner transforms supplies the implementation. It constructs discrete Wigner and Fourier-Wigner transforms on the cyclic group of odd prime order $p$, and shows that the Fourier-Wigner basis is extremely sparse (4,320 versus about 350,000 nonzero matrix elements at their test size). Gait is roughly periodic, so resampling one estimated stride cycle to a prime length (for example $p=31$ or $61$) makes the cyclic construction natural rather than an artificial restriction. For a strip signal $s$ on $\mathbb{Z}_p$, the descriptor is the ambiguity magnitude:
$$A_s(\tau,\nu) = \Big|\sum_{t\in\mathbb{Z}_p} s(t+\tau)\,\overline{s(t)}\,e^{-2\pi i \nu t/p}\Big|,$$
Here $\tau$ is a time lag, $\nu$ is a frequency offset, and the absolute value removes dependence on the cyclic start frame. Because $A_s$ is bilinear in $s$, it keeps relative-phase structure between stride harmonics, such as left-right swing asymmetry. A time-pooled spectrogram magnitude discards that structure.
The other quantum papers, including quantum clustering, state-discrimination limits, and semiclassical limits, do not enter the method. We do not claim any quantum advantage. The borrowed content is a representation and one theorem about what positivity costs.
Research Program
We build three things.
The signals. From each CASIA-B silhouette sequence we extract strip signals: for each of the horizontal strips GaitSet's pyramid already uses, the per-frame width and horizontal centroid. We estimate the stride period from the autocorrelation of the lowest strips, resample one or two cycles to prime length $p$, and normalize by cadence so that walking speed does not masquerade as identity.
The phase-space branch. A small convolutional network runs over the stacked strip ambiguity maps, and its embedding is concatenated with GaitSet's set-pooled embedding before the final metric-learning layer. The whole model is trained with GaitSet's own loss and schedule.
The protection test. We apply the Euclidean/Cosine locality-sensitive-hashing-plus-random-modulo scheme of Cancelable biometric schemes for Euclidean metric and Cosine metric to the combined gait embedding. That scheme hashes feature vectors with random projections and then applies a random many-to-one modulo map, needs no training, and was shown only on faces. We first use the hyperparameters that paper tuned for faces, then a small retuned grid. The point is to test whether one data-independent protection scheme generalizes across a static and a behavioral modality once the behavioral trait is turned into a stable vector.
The research file's overview asserts that such schemes fail catastrophically on behavioral traits, but the file reports no measurement of this. Our experiment provides one.
Baselines
Naive reference. The do-nothing reference is GaitSet itself (GaitSet), which reports rank-1 accuracy on CASIA-B of 95.0% (normal), 87.2% (bag) and 70.4% (coat), plus 87.1% on OU-MVLP. We retrain it ourselves under the identical protocol so that seed variance is known. Beating it means a coat-condition gain larger than three standard deviations across five seeds, with no loss on normal walking beyond seed noise. GaitSet is the strongest gait method in our file. It is also the right competitor because it embodies the opposite hypothesis: that timing is unnecessary.
Capacity control. We fuse GaitSet with a second, independently seeded GaitSet of matched parameter count. This separates gains from dynamics from generic ensembling gains.
Positive-smoothing ablation. This ablation isolates the key ingredient. It replaces the ambiguity maps with Gaussian-window spectrograms of the same strip signals, pooled to the same size and fed to an identical branch. Window width is tuned on the same validation split with the same number of trials as our branch.
Learned-dynamics competitor. A temporal convolution stack over GaitSet's per-frame features, with parameters and tuning trials matched to our branch, tests whether hand-specified phase space beats simply letting the network learn timing.
Protection baseline. On faces, the published reference is plain-template EER 0.125% rising to 0.154% with Euclidean LSH protection on CASIA-FaceV5 (Cancelable biometric schemes for Euclidean metric and Cosine metric). For protected gait templates, no reported result was found on Lacuna, so the comparison there is protected versus unprotected EER from our own models. Success means a relative EER increase on gait within a factor of two of the face-side relative increase, with face-tuned hyperparameters.
Why not just…?
Why not just add a temporal network to GaitSet and let it learn the dynamics? This is a real threat. The file contains no evidence that learned temporal models fail on coat-condition gait. The learned-dynamics competitor is run with matched parameters and tuning budget. If it equals or beats the phase-space branch, the conclusion is that timing matters but the Wigner representation adds nothing. We would report that and drop the representational claim. The phase-space branch could still win for two reasons. It sees cadence-normalized, start-invariant signals by construction. It also works from strip geometry rather than from appearance features that coats corrupt.
Why not just use a spectrogram, the standard time-frequency tool? The spectrogram with a Gaussian window is the Husimi function. The Ohst–Plávala theorem says any non-negative time-frequency picture must blur jointly in time and frequency. Whether that blur destroys identity-relevant harmonic phase in gait is an empirical question, and the positive-smoothing ablation answers it directly. If the spectrogram matches the signed branch, positivity costs nothing here, and the simpler tool should be preferred.
Why not just protect GaitSet's embedding directly, since set pooling already yields a fixed-length vector? This may well work. The claim that hashing schemes collapse on behavioral traits is not backed by a measurement in our file. We therefore run LSH-modulo on plain GaitSet embeddings as well as on the combined embedding. If plain GaitSet protects as gracefully, the protection argument for phase space disappears, and the contribution reduces to the recognition gain.
Proposed Experiments
All recognition experiments use CASIA-B with the GaitSet training and test split, and all numbers are averaged over five seeds. The primary metric is coat-condition rank-1 accuracy excluding identical views, and bag and normal conditions are reported alongside it. OU-MVLP tests whether any gain survives at more than 10,000 subjects.
Three decisive checks accompany the main table.
Frame-shuffle test. Shuffling frame order leaves GaitSet unchanged by construction. It should sharply degrade the phase-space branch's standalone accuracy. If it does not, the branch is not using dynamics and the story is wrong.
Lower-strip-only test. Restricting the phase-space branch to the lowest strips should retain most of its coat-condition gain, which would support the hypothesis that shin rhythm survives coats.
Short-clip test. At 7 frames GaitSet reports 82.5%. Our branch needs at least one stride cycle, so we expect it to help less or not at all on very short clips. We report this failure regime explicitly rather than hiding it.
For protection, we compute genuine and impostor EER on CASIA-B before and after LSH-modulo. This is done for plain GaitSet and for the combined embedding, with face-tuned and retuned hyperparameters. We also reproduce the face-side setting on CASIA-FaceV5 with an off-the-shelf face embedding, so that relative degradation across modalities is measured under one code path.
The project succeeds if the signed branch beats GaitSet, the ensemble control, and the spectrogram ablation on coats while failing the shuffle test as predicted. It is still informative if the spectrogram matches it, because that would show that for walking, as opposed to quantum states, positivity is nearly free.
References
- Symmetries and Wigner representations of operational theories
- Q-Flow: Generative Modeling for Differential Equations of Open Quantum Dynamics with Normalizing Flows
- On discrete Wigner transforms
- Cancelable biometric schemes for Euclidean metric and Cosine metric
- GaitSet: Regarding Gait as a Set for Cross-View Gait Recognition