ESP-Detection is a lightweight and ESP-optimized project based on Ultralytics YOLOv11, designed for real-time object detection on ESP series chips. It allows you effortlessly train a detection model for specific target and deploy the model on chips easily by ESP_DL.

ESP-Detection provides a series of ultra-lightweight models along with APIs that enables you to train custom detection models tailored to your specific use cases. The offered models are optimized for efficient deployment on ESP AI chips, like ESP32-P4 and ESP32-S3. The project also includes example applications such as cat detection, dog detection, and pedestrian detection.

• Competitive mAP: The single-class model ESPDet-Pico(0.36M parameters, 224 input size) achieves comparable mAP(0.5:0.95) to 80-class model YOLOv11n(2.6M parameters, 640 input size) on cat detection.
• Faster latency: Considering pre-processing and post-processing, espdet_pico achieves >14 FPS on ESP32-P4 and >6 FPS on ESP32-S3 when the input size is 224*224.
• Deploy friendly: A complete deployment solution is provided based on ESP-DL, offering users an effortless way to deploy custom models.

• mAP^val values are for single-model sing-scale on cat subset of COCO val2017 dataset.

• 2025/04/23: esp-detection 1.0.0 is public. Cat Detection is available.

conda create -n espdet python=3.8
conda activate espdet
pip install -r requirements.txt

• ESP-IDF is not required during model training, but is mandatory when running the quantized model on ESP chips. For setup instructions, please refer to ESP-IDF Programming Guide, and make sure to use ESP-IDF release/v5.3 or above.

Dataset format in esp-detection follows the YOLO detection dataset format. To convert your existing dataset from other formats (like COCO etc.) to YOLO format, please use JSON2YOLO tool by Ultralytics.

In esp-detection, we provide an all-in-one script espdet_run.py that streamlines the entire flow. With a single command, users can easily perform model training, export, quantization, and deployment. You can customize the espdet_run.sh shell script and execute the full pipeline with sh espdet_run.sh.

python espdet_run.py \
  --class_name mycat \
  --pretrained_path None \
  --dataset "cfg/datasets/coco_cat.yaml" \
  --size 224 \
  --target "esp32p4" \
  --calib_data "deploy/cat_calib" \
  --espdl "espdet_coco_224_224_mycat.espdl" \
  --img "espdet.jpg"

• MPS, CPU, Single-GPU and Multi-GPU training are supported in esp-detection. Please refer to Ultralytics YOLO Docs for more information. Specifically, you can set your own train settings according to Train Settings from Ultralytics.

Once Step 2 is completed, an example will be automatically generated. You can use it to perform model inference on ESP32-S3 or ESP32-P4. Similar to other esp-dl/examples, it enables customized model inference on a single input image and prints the results to the terminal.

You can run the example directly in the terminal with the following command:

cd path/to/custom_example
idf.py set-target esp32p4
idf.py flash monitor

• If replacing espdet.jpg with a custom image, ensure its width and height are correctly set in app_main.cpp.

Please submit an issue if you find any problems using our products, and we will reply as soon as possible.

• https://github.com/ultralytics/ultralytics