Detects7

Team: Code Enforcers

---

Summary

Detects7 is a web + API demo and research prototype for detecting seven safety-related object classes. It was trained on the Falcon synthetic dataset and deployed with a FastAPI backend and a Vite + React frontend. The project is designed to be simple to run locally and to serve as a foundation for further experiments.

Team

Name	GitHub	Role
Akash Kumar	XynaxDev	Project lead — directs model development, tunes experiments and hyperparameters, and manages API integration for deployment.
Lavnish	lavn1sh	Frontend & integration engineer — builds the web interface and connects the ML outputs to the UI and backend.
Nishtha	niishthaaaaaa	Model engineer — runs training experiments, prepares and curates the dataset, and iterates on model performance.
Himanshi	Himanshi1531	Research lead — surveys literature, recommends improvements, and helps shape experiment design.

App name: Detects7 — Team: Code Enforcers

Table: Code Enforcers — team roles and responsibilities.

Key results

Final metrics taken from ml/exp12/results.csv (final epoch = 150).

Metric	Value
Precision	0.91082
Recall	0.67942
mAP@50	0.75223
mAP@50:95	0.60106
Training epochs	150
Best checkpoint	ml/exp12/weights/best.pt (also exported in models/)

The evaluate.py script generates evaluation_summary.json and a confusion matrix for an official summary.

Repository layout (important files & folders)

Path	What it contains
backend/	FastAPI backend (see backend/app/ for main.py, model_loader.py, utils.py, config.py)
frontend/	Vite + React UI (src/ contains App.jsx, main.jsx, components, styles)
ml/	Training and evaluation code, dataset config and experiments (yolo_params.yaml, train_yolo.py, evaluate.py, predict_user.py, exp12/)
models/	Deployment artifacts (best_model_backup.pt, best_model_backup.onnx)
local_run.py	Convenience runner for local development
space/	Python virtual environment (optional)
README.md	This file

How to reproduce / common commands

Use the project virtual environment in space/ or create a fresh one.

1. Activate environment (PowerShell example)

& D:\detects7\space\Scripts\Activate.ps1
pip install -r requirements.txt
cd ml

2. Train (example using ml/exp12/args.yaml)

python train_yolo.py --data yolo_params.yaml --epochs 150 --imgsz 768 --batch 8

3. Evaluate (produces evaluation_summary.json)

python evaluate.py

4. Interactive prediction (image / video)

python predict_user.py

5. Backend (FastAPI) — example

cd backend
& ..\space\Scripts\Activate.ps1
pip install -r requirements.txt
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

6. Frontend

cd frontend
npm install
npm run dev
# then open http://localhost:5173

Models & artifacts

Artifact	Location
Best checkpoint	ml/exp12/weights/best.pt
Last checkpoint	ml/exp12/weights/last.pt
Exported deployment copies	models/best_model_backup.pt, models/best_model_backup.onnx

Tip: keep large model files out of Git history — use Git LFS or release assets for big binaries.

Design notes & decisions

• We initially explored RTDETR but switched to Ultralytics YOLOv8 because it gave better detection performance for this dataset.
• Training specifics (see ml/exp12/args.yaml): AdamW, lr0=0.001, imgsz=768, batch=8, epochs=150. Augmentations and other settings were tuned in the experiment.
• evaluate.py runs the Ultralytics evaluation suite and produces a confusion matrix and a JSON summary.

🗄️Dataset & acknowledgements

• Dataset: Falcon synthetic dataset — https://falcon.duality.ai/
• Model backbone: Ultralytics YOLOv8

---

Thank you 💌 — Detects7 team