TensorFlow 2 Object Detection API Tutorial

A maintained Windows and Conda workflow for fine-tuning TensorFlow 2 Object Detection API models on a Pascal VOC dataset. It replaces the pinned 2022 environment, global PYTHONPATH, and manual system-wide Protobuf installation.

The tracked models/ directory is an archived snapshot retained for historical reference. New work uses a fresh clone of the official TensorFlow Model Garden under third_party/models, installed inside the Conda environment.

1. Create the environment

TensorFlow is installed from its official pip wheels inside Conda. This project defaults to CPU execution; no CUDA, cuDNN, or GPU setup is required.

conda env create -f environment.yml
conda activate tf2-object-detection

2. Install Object Detection API

powershell -ExecutionPolicy Bypass -File scripts/setup_object_detection.ps1

The script clones the latest Model Garden, compiles protobuf definitions, installs the API into the active environment, and runs the official model-builder test. Use -SkipTests only when the installation has already been verified.

3. Prepare data

Place each image beside its Pascal VOC XML annotation in workspace/training_demo/images. Split and create records with explicit paths:

cd workspace/training_demo
python scripts/pre-process/partition_dataset.py --image-dir images --test-ratio 0.2 --copy-xml
python scripts/pre-process/generate_tfrecord.py --xml-dir images/train --image-dir images/train --labels-path annotations/label_map.pbtxt --output-path annotations/train.record
python scripts/pre-process/generate_tfrecord.py --xml-dir images/test --image-dir images/test --labels-path annotations/label_map.pbtxt --output-path annotations/test.record

4. Train and export

Download a model and matching pipeline.config from the official TF2 Detection Model Zoo. Set the model path, TFRecord paths, label-map path, batch size, class count, and fine-tune checkpoint in that configuration.

Then run the installed Model Garden commands from the repository root:

python third_party/models/research/object_detection/model_main_tf2.py --pipeline_config_path=workspace/training_demo/models/my_model/pipeline.config --model_dir=workspace/training_demo/models/my_model --alsologtostderr
python third_party/models/research/object_detection/exporter_main_v2.py --input_type=image_tensor --pipeline_config_path=workspace/training_demo/models/my_model/pipeline.config --trained_checkpoint_dir=workspace/training_demo/models/my_model --output_directory=workspace/training_demo/exported-models/my_model

The deployment helper remains in workspace/training_demo/scripts/process. All project scripts support --help.