EffiPed detects pedestrians across several fixed cameras and ranks which observations plausibly show the same person, so that a human can review them. For engineers and researchers evaluating compact multi-camera re-identification — and for anyone who wants to try that workflow in a browser, with nothing to install.

EffiPed

Multi-Camera Pedestrian Detection, Tracking & Re-Identification using Joint ConvNeXt V2 Architecture

By Aswanth Raj

EffiPed is a compact video-intelligence system that detects pedestrians, maintains camera-local tracks, and ranks cross-camera identity candidates for human review. Its React investigation console is available as a precomputed browser demo; the same workflow can connect to local FastAPI/CUDA inference when an authorized checkpoint is available.

Important

Ranked matches are reviewable appearance evidence, not proof of identity. The hosted experience is a precomputed, non-commercial research demonstration. Public model weights remain withheld while training-data redistribution terms are unresolved.

Try the identity review console

The hosted UI is an application console: a fixed top bar, one viewport, a thresholds panel that drops from the bar when it is asked for, and a status bar that reports the job. Four P-DESTRE session 12-11-2019_3 clips are already attached, as though you had uploaded them. Six workspaces, switchable with the number keys:

• Person Search (1), the workspace the console exists for. Pick a camera, pick anyone the pipeline found in it, and review the ranked candidates from the other cameras beside the full frame each one was cropped from.
• Detection (2), one frame with its stored detections redrawn at the current confidence and box-height floors.
• Tracking (3), a clip's tracked render beside a live table of the tracks in it.
• Cross Camera (4), the camera-pair link matrix, the strongest links at the current threshold, and the archived per-pair precision diagnostic.
• Sources (5), the attached clips and their specifications.
• Model (6), the loaded artifact, the settings this replay's index was built with, the benchmarks by protocol, and the boundaries.

A threshold moves the whole job rather than one panel: raising the confidence floor drops detections, which drops tracks, which drops links, and the status bar prints what each change cost.

cd apps/web
npm install
npm run dev

The controls are live, but the hosted build performs no inference: each Run replays a precomputed result. Full-frame views are redrawn in the browser by seeking the shipped clip to the appearance's timestamp and stroking its stored box, so no per-appearance scene images ship.

No synthetic browser boxes are drawn over the footage. The boxes baked into the replay videos are the annotations rendered by the original tracking pipeline.

Note

The person-search index was computed offline with the BoxJDE research checkpoint, because the EffiPed Tier-1 weights are withheld pending dataset rights review. The Person Search and Model Status panels both state this. Regenerate it with tools/build_person_search_fixture.py.

System

One ConvNeXt V2 feature hierarchy supports CenterNet-style detection and a 256-D part-aware identity descriptor. RoIAlign extracts a person feature map, four horizontal body strips retain local appearance, and Coordinate Attention fuses the visible evidence. BoT-SORT combines motion, overlap, and appearance for local temporal association; the gallery then ranks possible cross-camera matches for an analyst.

Evaluation	Result
P-DESTRE validation cross-camera Rank-1	62.8%
P-DESTRE test cross-camera Rank-1	61.3%
P-DESTRE validation / test detection mAP@0.5	90.74% / 88.4%
MOT17 val-half MOTA / IDF1 / HOTA	64.08 / 74.24 / 61.34
EffiPed Tier-1 footprint	7.78M · ≈18 full-pipeline FPS

Each value has a protocol label in RESULTS.md. The interactive replay is an application demonstration, not a benchmark run.

Architecture

The diagram is also available as an editable PowerPoint.

Repository map

src/effiped/          installable model, descriptors, tracking, runtime
apps/api/             FastAPI local-GPU service and job lifecycle
apps/web/             React/Vite demo workbench UI and hosted replay
configs/system/       active EffiPed and matched PartJDE configurations
research/results/     single source of truth for published evidence
research/report/      generated technical report
docs/architecture/    editable diagram source and web exports
docs/media/           optimized, attributed demonstration media
tools/ and tests/     validation, regression, and release checks

Run live inference locally

Python 3.11 and an NVIDIA GPU are recommended.

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
pip install -e ".[runtime]"
effiped-app

Place an authorized checkpoint in EFFIPED_WEIGHTS_DIR. When none is present, the API reports the model as unavailable without exposing a local filesystem path.

effiped-train --config configs/system/effiped-tier1.yaml
effiped-eval --config configs/system/effiped-tier1.yaml
effiped-demo

Variable	Purpose
EFFIPED_WEIGHTS_DIR	authorized local model artifacts
EFFIPED_RUNTIME_DIR	temporary uploads, crops, and job assets
EFFIPED_DEVICE	auto, cpu, cuda, or cuda:N
EFFIPED_MAX_UPLOAD_MB	per-video upload limit
EFFIPED_ALLOWED_ORIGINS	comma-separated CORS allowlist

Public API

• GET /api/health
• GET /api/models
• POST /api/person-search/jobs
• GET /api/person-search/jobs/{job_id} and /stream
• GET .../people, /detections, /tracks, and /matches
• POST .../search-by-example
• DELETE /api/person-search/jobs/{job_id}
• GET /api/assets/{asset_id}

Deleting a job removes uploaded video and generated assets.

Limitations

Written from what this repository can and cannot show, not from modesty.

• A ranked match is not an identification. Cross-camera similarity orders candidate appearance evidence for a person to review. The demo fixture makes the reason visible: non-matches score 0.997–0.998 against matches at 0.998–0.999, so the ranking is useful and the absolute score is not. There is no calibration and no decision threshold.
• The shipped service ranks cross-camera candidates by cosine similarity alone. PersonSearchManager.matches in apps/api/person_search_service.py compares the query descriptor against every other person, drops anyone from the query's own clip, sorts, and truncates. The cross_camera: policy in configs/system/effiped-tier1.yaml, which sets a match threshold, gallery size, temporal window, transition weight and a global identity cap, belongs to CrossCameraAssociator, and nothing outside tests/test_tracker.py reaches that class. The configured policy and the running behaviour are not the same thing, on the capability this project is named for.
• No published weights, so nothing here reproduces the numbers. Publication is on hold pending a dataset-rights review (DATA_LICENSES.md). pip install succeeds, effiped-app starts, and every model reports available: false. The reported results are attested by research/results/summary.json and defended against drift by tools/validate_results.py; they are not re-derivable from this repository alone.
• Every published number is a single measurement on one fold. No variance, interval, seed policy or significance test is reported for any value, and the FPS figure is approximate, from one device at one resolution.
• Evaluated only on P-DESTRE fold 0 and MOT17 val-half. Nothing here establishes behaviour for another site, population, camera network, or operating condition.
• No fairness or subgroup analysis, on a system that ranks people by appearance.
• Nothing measures the review loop the system exists for. There is no study of whether ranked candidates make a reviewer faster or more accurate.
• The hosted demo performs no inference. It replays archived output from the original PedestrianTracker application; the controls are live, but Run returns a stored result.
• The local API has no authentication, which is safe only because it binds 127.0.0.1. The container image binds 0.0.0.0, so publishing that port is a decision requiring its own review.
• Two-thirds of the Python is neither linted nor tested. train.py, loss.py and dataset.py are research code carried forward from the training workspace and are excluded from ruff; the inference path itself cannot be tested end to end without a checkpoint.

Research connections

The later BoxJDE Person Search repository isolates the full-person descriptor readout and documents its five-fold P-DESTRE ablation. It is linked as related research; its code and report are not duplicated here.

Licensing and responsible use

Original software is © 2026 Aswanth Raj and licensed under Apache-2.0. P-DESTRE-derived media under docs/media/pdestre/ is separately licensed as a CC BY-NC-SA 4.0 adaptation for this non-commercial demonstration. The asset manifest records the source, transformations, hash, purpose, and license for every derived asset.

No dataset, source video, person-level benchmark record, checkpoint, or runtime crop is included.

Model card · Data and weight-release audit · Third-party notices · P-DESTRE paper · CC BY-NC-SA 4.0