detrflow – End-to-End RT-DETR Object Detection Pipeline
Overview
detrflow is a fully end-to-end object detection pipeline built on RT-DETR (Real-Time Detection Transformer), covering everything from data loading and training to evaluation, serving, and deployment.
The project demonstrates a complete applied ML workflow — from pretrained baseline evaluation to a live inference API. A systematic knowledge distillation study on RT-DETR is in progress as a companion project.
Key Results
| Metric | Value |
|---|---|
| COCO val2017 AP (baseline) | 47.9 |
| Backbone | RT-DETR (ResNet-50) |
| Evaluation dataset | COCO val2017 (5,000 images) |
Features
- COCO val2017 Evaluation — Full benchmark against the standard detection suite with reproducible results
- FastAPI Inference API — REST endpoint for single-image and batch inference
- HuggingFace Space Deployment — Live demo hosted and publicly accessible
- Benchmark Scripts — Latency and throughput profiling utilities
- Model Card — Documented limitations, intended use, and evaluation results
Tech Stack
PyTorch · RT-DETR · HuggingFace Transformers · FastAPI · COCO API · Docker
Links
- 💻 GitHub Repository: github.com/umutonuryasar/detrflow
- 🤗 HuggingFace Space: huggingface.co/spaces/umutonuryasar/detrflow
Related Work
This project builds on the RT-DETR architecture and serves as the empirical foundation for an ongoing knowledge distillation study on RT-DETR. A related, completed study on KD dynamics across ResNet teacher-student pairs is available as an arXiv preprint (arXiv:2605.31191).