Traffic Sign Detection with TensorFlow Object Detection API

Instructions for training and evaluating networks from TensorFlow object detection model zoo for sign detection using the TT100K dataset. Tested on Ubuntu 16.04.5 LTS with Python 3.6, TensorFlow 1.12.0 and TensorFlow models v1.11.

AWS setup

If you are using AWS, follow the instructions here to set up an EC2 instance. Copy the TT100K dataset from S3 using

aws s3 sync s3://<path-to-tt100k>

If you're using a local GPU, skip this step.

Activate the in-built Conda environment for TensorFlow using

source activate <environment-name>

Refer the logs printed after you connect to the instance for the appropriate environment name.

Install TensorFlow object detection API

Note that AWS Deep Learning AMIs come with TensorFlow preinstalled but not the object detection API.

Clone the TensorFlow models repo from here.
```
mkdir tensorflow
git clone <url>
```
Follow the instructions [here] to complete the installation. You can use this script to install most of the dependencies. Make sure that you test the installation by running
```
python object_detection/builders/model_builder_test.py
```
from tensorflow/models/research. If you are getting errors, try reactivating the Conda environment.
If there are any CUDA linking issues while running the scripts, resolve them by running
```
sudo ldconfig /usr/local/cuda/lib64
```

Generate label map ProtoBuf

Generate the label map .pbtxt file by running

python code/generate_label_map.py -d <path-to-TT100K-dataset-root-directory> -o <path-to-output-pbtxt-file>

Create TF record files

Create training and test .record files by running

python create_tt100k_tf_record.py --data_dir=<path-to-TT100K-dataset-root-directory> --output_dir=<path-to-tfrecord-output-directory> --label_map_path=<path-to-label-map-pbtxt>

Here, output_dir is the path to the directory where you want to store the .record files which the Python script will generate.

Download pre-trained models

Create a folder named models
```
mkdir models
cd models
```

Download the model to be trained from the TensorFlow detection model zoo. For example, you can download SSD Mobilenet by running

wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2018_01_28.tar.gz
tar -xvzf ssd_mobilenet_v1_coco_2018_01_28.tar.gz

Change the number of classes and set appropriate paths in pipeline.config. Refer the file for more details.

Training

Train the model by running the following command from the tensorflow/models/research directory

python object_detection/model_main.py --pipeline_config_path <path-to-model-config-file> --model_dir <path-to-output-directory>

The above script does not print losses to the console. You can visualize training using TensorBoard by running
```
tensorboard --logdir=<path-to-output-directory>
```
where the output directory is the same as used in the first step.
If training stops for some reason or the loss is high and you want to train further, use the latest checkpoint in the config file and insert the following line after fine_tune_checkpoint:
```
load_all_detection_checkpoint_vars: true
```
This will reload the box predictor's weights along with those of the feature extractor.