I will demonstrate, how to create and use realtime object detection engine using YOLO and iOS. For network creation i use Ubuntu 19.04 with NVidia GPU. For iOS conversion and compilation i use Monterey and Xcode 13.1.

1. Get and compile darknet, i recommend AlexeyAB fork. Enable CUDA and OpenCV support.

2. Prepare image dataset. My network is for detection of SCRATCHES on 224x224 input. Refer to darknet docs if you need your own objects. Split images into scratch/positives and scratch/negatives. Positives must contain images with objects and txt files with boxes. Negatives must contain images without objects and empty txt files. You can use Yolo_mark.

3. Create yolo darknet model.

Use this method for devices with iOS >= 13. Currently script generates iOS15 MLProgram mlpackage, but can be easily modified for iOS13 and mlmodel. YOLOv4-TINY work well. Suddenly, large YOLOv4 mlpackage takes minutes to loading on every iOS example app launch. At least on iPhone12 with iOS15.0.1

1. coremltools is very sensitive to packages versions. This is why you need dedicated python anaconda environment. Install Anaconda from: https://repo.anaconda.com/archive/Anaconda3-5.3.1-MacOSX-x86_64.pkg.

(@junmcenroe reported good results with miniconda3-py37_4.10.3-MacOSX-x86_64.pkg on MacOS 10.15)

2. In Terminal enter conda environment (assuming anaconda installed to /anaconda3):

. /anaconda3/etc/profile.d/conda.sh
conda create -n coremltools-env python=3.7
conda activate coremltools-env
pip install yolov4==3.2.0
pip install opencv-python==4.5.4.60
pip install h5py==1.5.2
pip install coremltools==5.1.0
pip install keras==2.2.4
pip install tensorflow==2.5.0

PS: h5py==3.1.0 seems to be ok.

3. Prepare yolov4-tiny.cfg file (clear unsupported learning tags like subdivisions if any). Keep original yolov4-tiny.cfg for further trainings. Example:

sh ./prepare_cfg.sh yolov4-tiny.cfg yolov4-tiny_temp.cfg 

4. Use prepared yolov4-tiny_temp.cfg. Convert to iOS 15 MIL program target:

python ./convert_v4.py -n coco.names -c yolov4-tiny_temp.cfg -w yolov4-tiny.weights -m yolov4.mlpackage -l RGB

For large models not fitting iOS CoreML memory restrictions use iOS 14 neuralnetwork target:

python ./convert_v4_network.py -n coco.names -c yolov4-tiny_temp.cfg -w yolov4-tiny.weights -m yolov4.mlmodel -l RGB

5. Now I integrate anchors and names as spec for both mlpackage and mlmodel, so app code correctly loads such infromation from compiled model. Also different amount of yolo levels now authomatically detected and supported.

Use this method for unsupported devices with iOS < 13. Also see appropriate iOS App example.

1. Install Anaconda from: https://repo.anaconda.com/archive/Anaconda3-5.3.1-MacOSX-x86_64.pkg

2. yolo -> coreml:

conda create -n yolo2coreml python=3.6 anaconda
conda activate yolo2coreml
conda install tensorflow=1.14.0
conda install keras=2.3.1
conda install coremltools=4.1
python convert_v4_old.py yolov4.cfg yolov4.weights yolov4.mlmodel

3. You can use ios project as reference. Copy yolov4.mlmodel to project folder. Check anchors in yolov4.cfg and swift code. Change classes names and count, anchors, network size if you use your owns.

• YOLOv3-Tiny 224x224 (SCRATCH) network takes about 25 ms per detection on iPhone X.
• YOLOv4 old method 416x416 (COCO) network takes about 5 second per detection on iPhone 6.
• YOLOv4 608x608 (COCO) network takes about 10 seconds per detection on iPhone 12.
• YOLOv4-TINY 416x416 (COCO) network takes about 19 ms per detection on iPhone 12.

• YOLO
• AlexeyAB
• coremltools
• python-yolov4