Core ML Custom Object Detection

Setup

git clone https://github.com/bourdakos1/Core-ML-Custom-Object-Detection.git
cd Core-ML-Custom-Object-Detection

pip3 install -r requirements.txt

Usage

Convert Annotations

Convert your PASCAL VOC annotations to TensorFlow records.

annotations
├── labels
│   ├── label_map.pbtxt
│   ├── trainval.txt
│   ╰── xmls
│       ├── 1.xml
│       ├── 2.xml
│       ├── 3.xml
│       ╰── ...
╰── images
    ├── 1.jpg
    ├── 2.jpg
    ├── 3.jpg
    ╰── ...

python3 object_detection/create_tf_record.py

Train

python3 object_detection/train.py \
        --logtostderr \
        --train_dir=train \
        --pipeline_config_path=ssd.config

Create the TensorFlow Model

python3 object_detection/export_inference_graph.py \
        --input_type image_tensor \
        --pipeline_config_path train/pipeline.config \
        --trained_checkpoint_prefix train/model.ckpt-NUMBER \
        --output_directory output_inference_graph

Convert to Core ML

This script looks for the frozen_inference_graph.pb found in the output_inference_graph directory. It also looks for the label_map.pbtxt found in your annotations.

pip2 install tfcoreml tensorflow numpy protobuf
python2 core_ml_conversion/convert.py