YOLOv26n: Optimized QAT & Deployment on ESP32-P4

This repository provides an end-to-end workflow for Quantization Aware Training (QAT), Export, and Deployment of YOLOv26n (NMS-Free) on the ESP32-P4 SoC using the ESP-DL library.

By leveraging a custom dual-head architecture and direct regression (RegMax=1), this implementation achieves high-performance inference without the overhead of Distribution Focal Loss (DFL) or complex post-processing.

Tip

Accuracy Parity & Speed Leadership: This implementation achieves a 36.5% mAP matching the official ESP-DL YOLOv11nbut runs 30% faster thanks to hardware-specific SIMD kernels.

Benchmark Comparison

Architecture	Resolution	mAP (Accuracy)	Latency	Speedup
ESP-DL YOLOv11n	640px	36.0%	2.4s	1.0x
YOLO26n (Ours)	512px	36.5%	1.7s	1.3x (🚀 +30%)

---

Technical Case Study

Want to see how I achieved 1.7s latency on ESP32-P4?

⏱️ 5 Minute Read • 🧠 Deep Tech Analysis

---

Note

🚀 What's Next?

I am considering porting YOLO26n-Seg (Instance Segmentation) to the ESP32-P4 to enable pixel-level masks on the edge.

Support the Project: If you find this repository helpful, please consider giving it a Star ⭐.

Your support is my main motivation! If this project gains enough traction, I will prioritize building and open-sourcing the Segmentation pipeline next.

---

Bus Detection Analysis

Installation & Setup

This project requires the ESP-DL library to be located in the project root.

1. Clone Repository

git clone https://github.com/boumedinebillal/yolo26n_esp.git
cd yolo26n_esp

2. Install ESP-DL Library

The project relies on esp-dl for neural network kernels. Clone it into the root directory:

git clone https://github.com/espressif/esp-dl.git
cd esp-dl
git submodule update --init --recursive

Troubleshooting / Version Lock: While the latest version is recommended, if you encounter build errors, revert to the validated commit:

git checkout 7089b94a76206825bddba57a6385d46cc08c0a6b
git submodule update --init --recursive

Return to the project root:

cd ..

3. Install Python Dependencies

Ensure you have Python >3.9 installed.

pip install -r requirements.txt

Comparative Results (ESP32-P4)

Benchmark conducted on ESP32-P4 Engineering Sample.

Model Architecture	Configuration	Resolution	mAP (COCO)	Latency	Status
YOLO26n (Optimized)	Int8 QAT	512x512	36.5%	1.77s	Target
YOLO26n (High-Res)	Int8 QAT	640x640	38.5%	3.00s
YOLOv11n (Baseline)	Official ESP-DL	640x640	36.0%	2.75s

Key Features

• NMS-Free Architecture: Utilizes a One-to-One prediction head for direct inference, eliminating the need for Non-Maximum Suppression.
• Int8 Quantization: Full Quantization Aware Training (QAT) pipeline using esp-ppq for maximum efficiency on the P4 accelerator.
• Dynamic Resolution: Pre-configured support for both 512x512 and 640x640 resolutions in firmware.
• Optimized Firmware: C++ Inference engine (Yolo26Processor) optimized for ESP-DL's static graph execution.

Usage Pipeline

The entire QAT and Export workflow is orchestrated by YOLOv26_QAT_Workflow.ipynb.

1. Open the Notebook:

   jupyter notebook YOLOv26_QAT_Workflow.ipynb

2. Configure: Set your desired IMG_SZ (512 or 640) in the configuration cell.
3. Run Pipeline: Execute all cells to:
   • Export the PyTorch model to ONNX (with custom patches).
   • Quantize and Calibrate using esp-ppq.
   • Fine-tune (QAT) for optimal accuracy.
   • Export the final .espdl model artifacts.
4. Artifacts: The resulting models will be saved in output/size_512/ or output/size_640/.

Firmware Deployment

The ESP-IDF firmware project is located in yolo26n_esp32p4/.

Prerequisites

• ESP-IDF v5.5+ installed and sourced.

Deploying a New Model

After running the QAT pipeline:

1. Copy the generated .espdl file from output/ to yolo26n_esp32p4/main/models/.
2. Review yolo26n_esp32p4/README.md for specific build switches (e.g., selecting 512 vs 640 model).

Build & Flash

cd yolo26n_esp32p4
idf.py set-target esp32p4
idf.py build
idf.py flash monitor

Project Structure

.
├── yolo26n_esp32p4/      # ESP-IDF C++ Firmware Project
├── pipeline_source/      # Python Source (Trainer, Exporter, Config)
├── output/               # Generated Models & Checkpoints
├── YOLOv26_QAT_Workflow.ipynb # Main QAT Orchestrator
├── requirements.txt      # Python Dependencies
└── README.md             # Project Documentation

License

This project is licensed under the MIT License - see the LICENSE file for details.

Copyright (c) 2026 Billal Boumedine (https://github.com/boumedinebillal/yolo26n_esp)

Citation

If you use this work in your research or project, please cite:

@software{yolo26n_esp32p4,
  author = {Boumedine, Billal},
  title = {YOLO26n QAT & Deployment for ESP32-P4},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/boumedinebillal/yolo26n_esp}}
}

Contact

www.linkedin.com/in/boumedine-billal-a9a850239

Open an issue for bug reports or feature requests.

Pull requests are welcome!