Silhouette sequences
Normal walking sequences from CASIA-B, spanning 124 identities and 11 camera angles.
Undergraduate research · 2023 to 2025
Can the shape of motion become a signature?
We investigated how the way we compress a walking sequence into an image changes a neural network’s ability to recognise a person.
SUBJECT 057VIEW 090°
- 05
- Representations compared
- 02
- CNN architectures tested
- 11
- Camera viewpoints
- 124
- Walkers to tell apart
COMPUTER VISION ✦ BIOMETRICS ✦ DEEP LEARNING ✦ FEATURE ENGINEERING ✦ GAIT RECOGNITION ✦ PATTERN RECOGNITION ✦ OPTICAL FLOW ✦ NEURAL NETWORKS ✦ HUMAN IDENTIFICATION ✦ RESEARCH ✦
01 / THE QUESTION
Recognition begins
before the model.
A gait cycle may contain dozens of silhouette frames. A feature representation compresses that sequence into a form a classifier can learn - but every compression preserves some signals and discards others.
This project asked which representations carry the most useful identity information, and whether combining static shape with dynamic movement could improve recognition.
Research hypothesis
Representations that preserve both appearance and temporal change will outperform methods that isolate either signal.
02 / THE RESEARCH JOURNEY
Before we trained a model,
we mapped the field.
The project began as a structured literature review. Search results were progressively narrowed, relevant methods were logged in an extraction sheet, and evidence about strengths and weaknesses shaped what we eventually implemented.
LITERATURE SCREENINGClick a stage
STAGE 01 · DISCOVERY
6,952
A deliberately broad search
Search strings around gait, silhouettes, feature representation, and biometric recognition produced a wide pool across the databases recorded in the extraction sheet.
ScienceDirect5,348
IEEE Xplore837
Springer Link767
Source names and stage counts follow the workbook’s “Search Parameters” sheet. The written report names Scopus instead of Springer Link; the original records do not resolve that discrepancy.
KEPT
Inclusion criteria
- Published from 2014 onward
- Focused on silhouette feature representation
- Written in English
- Full text available
- Clear and relevant methodology
CUT
Reasons for exclusion
- Outside the gait-recognition question
- Non-silhouette or unrelated sensor focus
- Duplicate database records
- Insufficient methodological relevance
- Failed abstract or full-text review
72 was the core review set recorded by the original search funnel - not the total number of papers collected during the whole project.
THE EXTRACTION SHEET
Evidence became structured data.
For every candidate paper, the working sheet tracked its reference, title, objective, representation, publication year, and journal. Separate sheets counted reported strengths, weaknesses, and how often techniques appeared.
85 workbook entries84 unique titles95 archived PDFs7 analysis sheets
PAPER TITLEREPRESENTATIONJOURNAL / VENUE
No matching paper in the extraction-sheet inventory.
85 entries · Every titled row from the project workbook is included. One title appears twice with different recorded representations. Paper titles are shown without inline academic references for light exploration; full details remain in the extraction sheet and report.
WHAT THE REVIEW CHANGED
GEI was dominant and efficient, but repeatedly associated with information loss. GEnI captured variability; GFI captured motion but at higher computational cost. That tension led us first to compare representations under matched classifiers - and then to combine complementary cues through DEWGI and the GaitSTAR adaptation.
Continue to the experiments ↓03 / EXPERIMENTAL PIPELINE
From a walk to a prediction.
Feature extraction
Each cycle was encoded as GEI, GEnI, GFI, MSI, or the proposed DEWGI representation.
Classification
Matched datasets were trained and evaluated with DenseNet and ResNet architectures.
Comparison
Final test accuracy and cross-entropy loss exposed the effect of representation choice.
Protocol scope: this project used CASIA-B normal walking conditions only. Reported metrics should be read within that experimental setup.
04 / INTERACTIVE FEATURE LAB
One walk.
Five interpretations.
Select a method to see how the same subject and view become a different machine-readable signal.
SAMPLE #057 - NM01 - 090°
GEIAPPEARANCE-BASED
Gait Energy Image
Averages every silhouette in the gait cycle. Stable body regions remain bright; moving limbs become softer traces.
- Captures
- Shape + averaged motion
- DenseNet result
- 88.41%
- ResNet result
- 81.42%
WHY TWO CLASSIFIERS?
Testing every representation with two different CNN families helped us ask whether performance came from the input representation or depended on one particular architecture.
Reuse what earlier layers learned.
DenseNet connects each layer to later layers, encouraging feature reuse and stronger gradient flow. In our matched experiments it achieved the higher accuracy and lower cross-entropy loss for all five representations.
BEST PAIR · DEWGI 92.34% · LOSS 0.27Learn corrections through shortcuts.
ResNet uses residual connections so deeper blocks learn adjustments to an identity path. It provided a second architectural test of each representation and also ranked DEWGI first.
BEST PAIR · DEWGI 88.17% · LOSS 0.5105 / VIEWPOINT CHALLENGE
The same person looks different from every angle.
Viewpoint variation is one of gait recognition’s defining challenges. Drag the control to inspect the project’s DEWGI output across all eleven CASIA-B camera positions.
090°000°090°180°
06 / ENHANCED GAITSTAR
CONTRIBUTION 02
We did not stop at representation comparison.
GaitSTAR combines spatial appearance with temporal motion and uses attention to emphasise discriminative evidence. We carried that premise into a project-specific feature pipeline and a deeper recognition network.
Instead of treating the method as a black-box baseline, we reworked preprocessing, feature fusion, classifier depth, supervision, augmentation, and optimisation.
FUSED REPRESENTATION · SUBJECT 057 · 090°
CNN appearance features + dense optical flow + temporal weighting
CNN appearance features + dense optical flow + temporal weighting
017×7 stemBroad spatial context
→
02Residual stages3 · 4 · 6 · 3 blocks
CHANNEL ATTENTION × 4
→
03512-D featureGlobal pool + reduction
→
04Dual headsMain + auxiliary supervision
WHAT CHANGED · 01
A fused sequence descriptor
CLAHE and bilateral filtering prepare each frame. A small CNN extracts appearance cues while Farneback optical flow captures magnitude and direction. Temporal weights pool both streams into one feature image.
CLAHE contrastEdge-preserving filterDense optical flowTemporal pooling
BEST RECORDED TEST ACCURACY98.712%
124-way classification under the project’s normal-condition split. An earlier classifier iteration recorded 94.911%; because architecture, optimisation, augmentation, and epoch count changed together, that comparison is a development milestone - not a component-by-component ablation.
WHY THERE IS NO LEADERBOARD HERE
The final report set this number beside published CASIA-B results (GaitGraph 87.7%, Gait-D 91.6%, GaitPart 96.2%, GaitSTAR 97.4%). Those figures come from protocols that test on people the model never saw in training, across camera angles. Our split was a random 80/20 split of images, so every test walker also appeared in training, and the same held-out set guided model selection.
That makes 98.712% a closed-set result on our own split. It is not comparable to those benchmarks, so we don't rank against them. Repeating the experiment under the standard subject-disjoint protocol is the obvious next step.
WHY WE FRAME THIS CAREFULLY
The 98.712% value is a saved notebook result, not a peer-reviewed claim. Some analysis cells later in the notebook use simulated confidence values for visualisation; those plots are not presented here as experimental evidence. The showcase separates recorded classification accuracy from illustrative analysis.
07 / SFRT RESULTS
The representation mattered.
08 / REPRODUCIBILITY RECORD
What the notebooks let us reconstruct.
This record comes from executable cells, saved outputs, and notebook metadata - not retrospective estimates. Open each panel for the exact experimental choices and known limitations.
01Data and split＋
GEI, GEnI, GFI, and DEWGI contain 8,155 derived images; MSI and Enhanced GaitSTAR record 8,154. Images were grayscale at 240×240 with 124 identity labels.
- 80/20 row-level random split
random_state=42for splitting- 6,524/1,631 samples for most methods
- 6,523/1,631 for MSI and GaitSTAR
- Training-mean centering applied to test data
02DenseNet recipe＋
All five representations used the same DenseNet training function for the matched comparison.
- Adam, learning rate 0.0001
- Categorical cross-entropy
- Batch size 32; maximum 50 epochs
- Rotation ±10°, width/height shifts 0.1, horizontal flips
- Early stopping on holdout accuracy, patience 10
03ResNet recipe＋
The ResNet experiments used a residual image classifier with checkpoints and scheduled learning rates.
- Adam, initial learning rate 0.001
- Categorical cross-entropy
- Batch size 32; 100 epochs
- Learning-rate schedule plus ReduceLROnPlateau
- Best checkpoint selected by holdout accuracy
04Enhanced GaitSTAR recipe＋
The final PyTorch model used the enhanced residual-attention architecture documented above.
- Batch size 32; 150 epochs
- AdamW, learning rate 0.001, weight decay 0.05
- Label smoothing 0.1
- Cosine OneCycle learning-rate schedule
- Auxiliary loss weight 0.4; gradient norm capped at 1.0
05Environment and limitations＋
The notebooks record Python 3.10.14, Kaggle GPU runtimes, CUDA execution, and Tesla T4 acceleration for most DenseNet/ResNet runs. Kaggle image IDs range from 30775 to 30787.
- Exact TensorFlow, PyTorch, and CUDA package versions were not printed
- No global NumPy/TensorFlow/PyTorch training seeds were set
- The holdout set was used for monitoring and final evaluation
- Enhanced GaitSTAR saved the epoch with the best holdout/test accuracy
- This means results are reproducible in structure, but not guaranteed bit-for-bit
09 / WHAT I LEARNED
“The input is not just preparation for the model. It is part of the model’s argument about the world.”
01
Feature design changes outcomes
The same classifier shifted materially when the sequence encoding changed.
02
Fusion was effective
DEWGI’s mix of static and dynamic cues was the most reliable of the five representations tested.
03
Scope shapes every claim
Strong normal-condition results do not automatically imply robustness to clothing, bags, or unseen settings.
10 / LIMITS AND NEXT QUESTIONS
What this does
not show yet.
The comparison between representations is fair, because every input went through the same classifiers on the same split. The absolute accuracies are optimistic, and the reasons are worth stating plainly.
- Same walkers in training and test. The split was by image, not by person, so the models were never asked to recognise someone new.
- No separate validation set. Early stopping and checkpoint selection used the same held-out images that produced the final scores.
- Normal walking only. Coats, bags and other CASIA-B conditions were not tested, and neither was any other dataset.
- Single runs. No seeds were fixed and nothing was repeated, so there are no error bars; the Enhanced GaitSTAR changes were not ablated one by one.
The questions I would take further: does a representation's advantage survive a subject-disjoint, cross-view test; which representations stay stable when clothing, carried objects or the camera change; and can a network learn the compression itself instead of being handed a hand-designed one.
Research interests
Computer vision and visual representation learning: how what a model is given to look at shapes what it can learn, and how to make that hold up when the viewpoint, the person or the setting changes.
11 / CONTRIBUTIONS
A collaborative report.
A codebase I led.
Clear authorship matters. The project combined complementary strengths in implementation, analysis, and academic reading.
Code, methodology, and experiments
- Authored most of the project code and all code published in this repository
- Implemented feature extraction and classification experiments
- Wrote Chapter 3: Methodology
- Wrote Chapter 4: Results and Discussion
- Co-wrote Chapter 5: Conclusion with Jake
- Contributed to literature collection, screening, filtering, and the wider research process
Literature, data collection, and research writing
- Wrote Chapter 1: Introduction
- Wrote Chapter 2: Literature Review
- Contributed substantially to data collection, literature screening, filtering, and the wider research process
- Co-wrote Chapter 5: Conclusion with Dillion
Dr Aziz Abdullah Dwumfuor
Supervised the final-year project at the Department of Computer Science, University of Ghana.
PROJECT
Impact of Silhouette Feature Representation on Gait Analysis
An unpublished final-year undergraduate research project carried out at the Department of Computer Science, University of Ghana. The project documentation was submitted in September 2024 and the work continued into January 2025.
- Researchers
- Dillion Amartey
Jake Sompa Asamoah - Supervisor
- Dr Aziz Abdullah Dwumfuor
- Tools
- Python · TensorFlow · PyTorch
OpenCV · Kaggle - Contact
- Email
LinkedIn ↗
Portfolio ↗