Apppublic

SasankaBandara/rtdetr-esod-research

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

RT-DETR Research Prototype: Small Object Detection Enhancement

A full-stack web application demonstrating enhancements to the RT-DETR architecture, specifically targeting Small Object Detection (SOD) in drone imagery (VisDrone2019 dataset).

This repository contains the deployment code for serving 4 variants of the RT-DETR ResNet-18 model, complete with a glassmorphic user interface for real-time inference and metric analysis.

✨ Key Features

—Multi-Model Interface: Seamlessly switch between the Baseline RT-DETR and 3 improved research variants.
—Real-Time Inference: Drag-and-drop image upload with instant bounding box generation using ONNX Runtime.
—Detailed Metrics Dashboard: Interactive tabs displaying:
—Class Metrics: Per-class AP scores (e.g., Pedestrian, Bicycle).
—Layer Analysis: Parameter distribution between Backbone and Decoder.
—Inference Stats: Latency and FPS benchmarks.
—Premium UI: Fully responsive, dark-themed glassmorphism design.

🚀 The Models

This application compares four model variants trained on the VisDrone2019 dataset:

Model ID	Variant Name	Key Innovation	mAP (50)	FPS (GPU)
`base_rtdetr`	Baseline	Standard RT-DETR-R18	Baseline	N/A
`p2_p3_fusion`	P2-P3 Fusion	High-resolution feature fusion	41.18%	~60
`query_imp`	Query IMP	Scale-aware query initialization	36.74%	~58
`aware_loss`	Aware Loss	Difficulty-aware loss reweighting	18.85%	~60

🛠️ Tech Stack

—Frontend: Vanilla HTML5, CSS3 (Glassmorphism), JavaScript (ES6).
—Backend: Python FastAPI.
—ML Inference: ONNX Runtime (GPU/CPU).
—Containerization: Docker.

💻 Quick Start

Option A: Docker (Recommended)

You can build and run the entire application with a single command:

bash

docker build -t rtdetr-app .
docker run -p 8000:8000 rtdetr-app

Access the app at http://localhost:8000.

Option B: Local Development

1.Clone the repository (Ensure you use Git LFS for models):

bash

    git clone https://github.com/YourUsername/rt-detr-webapp.git
    cd rt-detr-webapp

1.Install Dependencies:

bash

    pip install -r backend/requirements.txt

1.Run Server:

bash

    python run.py

1.Open in Browser: Navigate to http://localhost:8000/frontend/index.html (or http://localhost:8000 if using the unified serving method).

📂 Project Structure

├── backend/
│   ├── models/          # .onnx model files (tracked by LFS)
│   ├── main.py          # FastAPI application
│   ├── model_utils.py   # Preprocessing & Inference logic
│   └── models_config.json # Auto-generated metrics DB
├── frontend/
│   ├── index.html       # Main UI
│   ├── style.css        # Premium styling
│   └── script.js        # UI Logic & API calls
├── Dockerfile           # Deployment configuration
├── run.py               # Local execution script
└── requirements.txt     # Python dependencies

⚠️ Note on Large Files

This repository uses Git LFS to store the .onnx model files (~300MB each). Ensure you have Git LFS installed before cloning.

Created for Research Prototype Demonstration.