We recommend creating a python 3.6+ virtual environment for this project. You can use pyenv-virtualenv to do so.

Install these Tensorflow versions in the activated environment.

tensorboard==1.14.0
tensorboard-plugin-wit==1.6.0.post3
tensorflow-estimator==1.14.0
tensorflow-gpu==1.14.0

• Download the necessary scripts with git clone https://github.com/tensorflow/models.git
• Install TensorFlow Object Detection API by strictly following these instructions. Once you've successfully run python object_detection/builders/model_builder_test.py you are ready for the next step.
• To access the necessary utility scripts, you'll need to run all the following commands from the models/research/object_detection directory from the cloned repo. From here on we will refer the TensorFlow Object Detection directory models/research/object_detection/ as the TOD directory.

You could also work from this codebase as a stable implementation with the above listed TF library versions. Just ensure that repo folder is set as models/research/object_detection/.

Tensorflow API supports a variety of file formats. The TFRecord file format is a simple record-oriented binary format that many TensorFlow applications use. We have example code in this repo which converts the labels.npz file to a TFRecords file:

• Copy utils_convert_tfrecords.py from this repo to the TOD directory, .
• Your $folder will be the data path containing your labels.npz file and tiles.
• From the TOD directory run:

python3 utils_convert_tfrecords.py    \
        --label_input=$folder/labels.npz   \
        --data_dir=tf_records   \
        --tiles_dir=$folder/tiles    \
        --pbtxt=classes.pbtxt

This will create train.record, val.record and test.record files in a folder called tf_records in the TOD directory. Each record file contains different and non-overlapping partitions of the data (86,7,7 percents, respectively).

Now we're ready to set up the model architecture. For this walkthrough, we'll download a pre-trained model from the TensorFlow model zoo. We'll demonstrate using ssd_resnet_101_fpn_oidv4 (download link):

• Download the model, unzip, and move the folder to the TOD directory
• Create a new folder training in the TOD directory.
• Copy a model configuration file to the training directory.
• Copy a class definitions file to the data directory.

Now your current directory should be models/research/object_detection/ and in addition to the files included in that repo originally, your folder structure should look like this:

models/research/object_detection/
├── ssd_resnet101_v1_fpn_multilabel/
├── training/
│   └── ssd_resnet101_v1_fpn_marine_debris.config
├── data/
│   ├── train.record
│   ├── val.record
│   ├── test.record
│   ├── marine_debris.pbtxt
└───

You are now ready to train the model. From the models/research/ directory, run:

#!/usr/bin/env bash
pyenv activate tf114_od
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
cd object_detection
export CUDA_VISIBLE_DEVICES=0
python model_main.py --alsologtostderr --model_dir=training/ --pipeline_config_path=training/ssd_resnet101_v1_fpn_multilabel.config

The model checkpoints and outputs for this task will save in the training folder.

Using this script, create the marine debris detection model inference graph with:

python export_inference_graph.py --input_type image_tensor \
              --pipeline_config_path training/ssd_resnet101_v1_fpn_multilabel.config \
              --trained_checkpoint_prefix training/model.ckpt-500000 \
              --output_directory model_50k

We can visualize this graph using tensorboard:

tensorboard --logdir='training'

Go to http://127.0.0.1:6006/ in your web browser and you will see:

Now let's run the model over our test tiles to predict where marine debris patches are. Copy this script to the TOD directory then run:

python tf_od_predict_image_aug_to_geo_corrected.py --model_name=model_50k \
                        --path_to_label=data/marine_debris.pbtxt \
                        --test_image_path=path/to/test/image/tiles

This code will read through all your test images in path/to/test/image/tiles folder and output the final prediction into the same folder. You will find new images in test_image_path with the _test suffixed to the end of the file basenames. The are images with the predicted bounding boxes and confidence scores plotted on top. As well, you will find a multipolygon geojson of predicted bounding boxes in the test_image_path.

Option for flags:

export base_dir=models/research/object_detection
export EXPORT_DIR=models/research/object_detection/model_50k
python3 ${base_dir}/tf_od_predict_image_aug_to_geo_corrected.py --model_name=${EXPORT_DIR} --path_to_label=${base_dir}/marine_debris.pbtxt --test_image_path=${base_dir}/test/

Detections geo-registered and vectorized to GeoJSON format:

You can use the code in this folder to compute standard evaluation metrics with your model. Runtime and background instructions live here.