ODModel

The object detection model for ARMAPS. This model is used to create a model for the objects in the sorounding that will be used to validate the AR output.

Getting Started

Tutorial - https://towardsdatascience.com/creating-your-own-object-detector-ad69dda69c85

TensorFlow

1. Download tensorflow

git clone https://github.com/tensorflow/models.git

2. Install tensorflow First modify the dockerfile at research/object_detection/dockerfiles/tf1/, by adding below at line 15

libgl1-mesa-dev

and add below at line 41

RUN pip install jupyterlab
RUN pip install opencv-python

cd models
docker build -f research/object_detection/dockerfiles/tf1/Dockerfile -t od .

3. Run tensorflow

cd armaps-od
docker-compose up -d

4. Enter into the tensorflow environment

docker exec -it armaps-od_tensorflow_1 /bin/bash

5. Test tensorflow to ensure its working correctly

python object_detection/builders/model_builder_tf1_test.py

Label Img

1. Download labelImg

git clone https://github.com/tzutalin/labelImg.git

2. Install on linux. To install on another os follow the instructions https://github.com/tzutalin/labelImg#installation

cd labelImg
sudo apt-get install pyqt5-dev-tools
sudo pip3 install -r requirements/requirements-linux-python3.txt
make qt5py3

3. Run label img

python3 labelImg.py

Python environemnt locally

1. Setup virtual environment

cd armaps-od
python -m venv venv

2. Activate virtual environment

source venv/bin/activate (linux)

venv\Scripts\activate (Windows)

3. Intsall Requirments

pip install -r requirements.txt 

If a new python package is added refresh the requirments filr with

pip freeze > requirements.txt

Run Jupyter

jupyter notebook --port=8888 --no-browser --ip=0.0.0.0 --allow-root

Steps to train and test model (Perform all these comands in the tensorflow docker container)

1. Transforming image sizes**

python transform_image_resolution.py -d images/ -s 800 600

2. Label the images using labelImg

3. Transform xml to csv

python xml_to_csv.py

4. Generate Tensor flow record**

python generate_tfrecord.py --csv_input=images/train_labels.csv --image_dir=images/train --output_path=records/train.record
python generate_tfrecord.py --csv_input=images/test_labels.csv --image_dir=images/test --output_path=records/test.record

5. Train the model

python model_main.py --logtostderr --model_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config

6. In another terminal start enter the comands to view the tensorflow logs

docker exec -it armaps-od_tensorflow_1 /bin/bash
cd object_detection
tensorboard --logdir=training

7. Export the inference graph for the results (change XXX to the highest number in the training fodler)

python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/faster_rcnn_inception_v2_pets.config --trained_checkpoint_prefix training/model.ckpt-XXXX --output_directory inference_graph

8. Test the model
9. Convert to Js compatible format Run below inside of the container

pip install tensorflowjs[wizard]
cd object_detection/inference_graph
tensorflowjs_wizard

tensorflowjs_converter
--input_format=tf_saved_model
--output_node_names='MobilenetV1/Predictions/Reshape_1'
--saved_model_tags=serve
saved_model
web_model2