StepUP 2026: Biometric Footstep Verification

Verifying identity from paired footstep pressure recordings across changes in footwear and walking speed.

Method | Reproduction | Final submission | Paper

Code for the ArogyaPandit Research Team submission to the 2nd International StepUP Competition for Biometric Footstep Recognition at IJCB 2026.

The organizers' final validated standings place the team 1st with 8.00% EER, as reported in the competition summary paper. Our final submission is SUBMIT_TOMORROW_ms400_weight_115.zip.

Research Question

Can a system verify an unfamiliar person's identity from a walking stride when enrollment and query samples have different shoes or walking speeds?

The input is a left/right pair of time-varying pressure recordings and a claimed identity. The output is a similarity score against that identity's reference recordings. A higher score indicates stronger support for the claim. The task requires generalization to people absent from training, with limited enrollment data and distractor identities in the probe set.

Footwear changes the contact pattern recorded by the mat, so a model can learn shoe appearance instead of identity. Our approach combines a learned stride representation with cohort-based normalization and score fusion to address this mismatch.

Pipeline

Left pressure  --> shared CNN --+
                                +--> concatenate --> FC fusion --> 512-D stride embedding
Right pressure --> shared CNN --+                                     |
                                                                L2 normalize
                                                                     |
                                                      probe/reference cosine scores
                                                                     |
                                                      training-cohort top-k S-norm
                                                                     |
                                                      reference-score aggregation
                                                                     |
                                                      weighted empirical-rank fusion
                                                                     |
                                                      claim verification score

Team

• Anant Gupta
• Gourav Gupta

Affiliation: Arogyapandit Private Limited, India

Method Summary

The system learns stride-level biometric embeddings from paired left and right pressure footsteps. Each foot is processed by a shared 3D CNN backbone. The two foot embeddings are concatenated, projected to a 512-dimensional stride embedding, and L2-normalized for verification.

Training used the provided competition training data only. Participants 1-140 were used for training/cohort generation and participants 141-150 were reserved for subject-disjoint validation. The documented approach uses Circle Loss (m=0.25, gamma=256) and prototype scoring against each claimed user's reference templates.

At inference, cosine claim scores were normalized with symmetric score normalization (S-norm) using cohort embeddings from the training participants. The final submitted file, SUBMIT_TOMORROW_ms400_weight_115.zip, used weighted empirical-rank fusion of two S-norm score lists:

final_score =
  0.885 * rank(primary_circle_snorm_ensemble)
+ 0.115 * rank(secondary_top400_mean_snorm_model)

No external biometric datasets or pretrained backbones were used.

The generator exports prototype-cosine and nearest-reference baselines separately. Its S-norm mean candidates average normalized similarities over individual references; that is different from cosine to a normalized prototype. Right-foot volumes are mirrored along the width axis, resized to 64 x 64 x 32 in (time, height, width) order, and min-max normalized per foot.

Results and Interpretation

The summary paper reports 8.00% EER and 42.88% FNMR at FMR = 1%. At the submitted threshold, FNMR is 0.16%, FMR is 32.80%, and balanced accuracy is 83.52%. These are different operating points; EER is not classification accuracy. This is a competition result, not evidence of deployment readiness.

Aggregate leaderboard feedback was available during development. The result should therefore be read as competition performance rather than an untouched test-set estimate. Empirical-rank fusion depends on the supplied batch of claims and is not a calibrated per-probe probability.

Release Scope

This release contains a baseline runner, scoring utilities, and the final fusion rule. Original winning checkpoints, complete training configurations, the construction of the primary ensemble, and original component scores are not included. The commands below do not reproduce the 8.00% result by themselves.

Historical checks recorded exact agreement with all 10,000 final submission scores given the two original inputs. Those inputs are absent from this checkout, so that comparison cannot currently be rerun. See reproduction details for the supported workflow and limits.

Repository Layout

code/stepup_runner.py          Train/infer/check CLI for stride embeddings
code/make_v2_submissions.py    Builds prototype, nearest, and S-norm submission candidates
code/fuse_rank_scores.py       Recreates SUBMIT_TOMORROW_ms400_weight_115.zip score fusion
code/snorm.py                  S-norm helper
docs/approach.md               Short paper-style approach description
docs/method.md                 Model, scoring equations, and limitations
docs/reproduction.md           Commands and checkpoint compatibility
CITATION.cff                   Citation metadata for the summary paper

Competition data, checkpoints, embeddings, and submitted score files are not included.

Expected Dataset Layout

data/competition/
  1 - Training/
  2 - Reference/
    reference_data.npz
    reference_metadata.csv
  3 - Probe/
    probe_data.npz
    probe_metadata.csv

Setup

Use Python 3.10 or newer.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Install the PyTorch build that matches your CUDA environment if the default package is not appropriate for your GPU.

Basic Usage

Check the dataset:

python code/stepup_runner.py check --dataset-root data/competition

Train a Circle Loss baseline:

python code/stepup_runner.py train \
  --dataset-root data/competition \
  --output-dir runs/circle_baseline \
  --loss circle \
  --epochs 10 \
  --batch-size 128 \
  --num-workers 0

Generate S-norm submission candidates:

python code/make_v2_submissions.py \
  --dataset-root data/competition \
  --checkpoint runs/circle_baseline/best.pt \
  --output-dir runs/circle_baseline/submissions \
  --top-k 100 200 400 \
  --num-workers 0

Fuse two S-norm score lists with the final SUBMIT_TOMORROW_ms400_weight_115.zip weighting:

python code/fuse_rank_scores.py \
  --primary path/to/primary_submission.zip \
  --secondary path/to/secondary_submission.zip \
  --secondary-weight 0.115 \
  --output runs/final_rank_fusion.zip

The two input lists must refer to the same claims in the same order. The original lists must be supplied separately. Start with zero data-loader workers; increase after checking throughput and memory on your machine.

Paper and Citation

The team contribution appears in The 2nd International StepUP Competition for Biometric Footstep Recognition: From Steps to Strides, by Robyn Larracy, Anant Gupta, Gourav Gupta, and colleagues. The arXiv record states acceptance to IJCB 2026. Use CITATION.cff to cite the paper.

Acknowledgments and Provenance

The runner adapts the public notebook workflow from the UNB StepUP organizer starter kit. We acknowledge the dataset creators, competition organizers, and the PyTorch Metric Learning implementation of Circle Loss. Git history is preserved from the earlier team release.

No new license grant is added by this documentation update. A project license is not currently included; consult the authors and upstream terms before redistribution. Competition data is obtained separately under its applicable terms.