Logo Classification Capstone Project

Installation Instructions

Clone the git repository of the Logo Capstone Project from Github.
Install Anaconda
Clone Tensorflow's Object Detection API from Github.
Either execute

# From Logo Capstone Project directory
$ conda create --name mypy3

and follow the installation instructions, or install the required packages with

# From Logo Capstone Project directory
$ conda create --name mypy3 --file requirements-mypy3.txt

Getting the Logos in the Wild dataset

Ask for permission and then download the Logos in the Wild dataset to the Logo Capstone Project directory.
Download the superset QMUL-OpenLogo Dataset (4.7 GB) that contains all available JPEGs to the Logos In the Wild dataset.
After changing in Logo Capstone Project/move_JPEG_files.sh the variable "oldpath" to the absolute path of your LogosInTheWild-v2/data-directory and "newpath" to the absolute path of your openlogo/JPEGImages-directory, enter the following:

# From Logo Capstone Project directory
$ ./move_JPEG_files.sh

In order to remove XML files without corresponding JPEG images from the Logos in the Wild dataset and du adjust brand names, remove the 0samples folder from the LogosInTheWild-v2/data directory, and execute in a separate Conda environment with Python 2.7 and opencv-python

# From LogosInTheWild-v2/scripts directory
$ python create_clean_dataset.py --roi --in ../data --out ../cleaned-data

This outputs that 9,428 images and 821 brands were processed, while 1,330 JPEG files were unavailable.

Custom Logo Classification CNNs in Keras

Open Logo Capstone Project/logo_classification.ipynb in a jupyter notebook, adjust the paths and execute the cells.

Creating TFRecord files for Tensorflow's Object Dectection API

For this necessary conversion of our dataset to Tensorflow's TFRecord file format, copy our adaption of Tensorflow's create_pascal_tf_record.py script create_and_analyze_pascal_tf_record.py from the Logo Capstone Project to the LogosInTheWild-v2 directoy.
In order to convert the Logos in the Wild dataset to TFRecord files, we run

# From LogosInTheWild-v2 directory
$ python create_and_analyze_pascal_tf_record.py --data_dir=./data/voc_format --label_map_path=./data/pascal_label_map.pbtxt --output_path=./data/

This converts 6,034 images with annotations into the 10 training TFRecord files \path{pascal_train.record-0000i-of-00010}, and 1,508 images with annotations into the 10 validation TFRecord files \path{pascal_val.record-0000i-of-00010} for i=0,...,9. Further, in LogosInTheWild-v2/data this created \path{test_images.txt} containing the absolute path of the 1,886 images in the test set, and pascal_label_map.pbtxt containing 821 entries like the following:

item {
  id: 262
  name: 'starbucks-text'
}

Training a custom Faster R-CNN with Tensorflow's Object Dectection API

Make sure to move the respective files (from the Capstone Project Repository) so that the directory structure is as recommended in Tensorflow's running_locally.md:

"data" directory
- pascal_label_map.pbtxt
- 10 files pascal_train.record-0000i-of-00010 for i = 0,...,9
- 10 files pascal_val.record-0000i-of-00010 i = 0,...,9
"models" directory
- "model" directory
  - faster_rcnn_inception_logos-locally-on-ubuntu.config
  - "train" directory
  - "eval" directory,

Train the Faster R-CNN on our dataset by running the following Python script:

# From tensorflow/models/research directory
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
python object_detection/model_main.py \
    --model_dir=PATH_TO/LogosInTheWild-v2/models/model/ \
    --pipeline_config_path=PATH_TO/LogosInTheWild-v2/models/model/faster_rcnn_inception_logos-locally-on-ubuntu.config \
    --num_train_steps=50000 --alsologtostderr

Monitor statistics with "tensorboard --logdir=PATH_TO/LogosInTheWild-v2/models/model/

Export the trained Faster R-CNN and test Logo detection

# From tensorflow/models/research/ directory
CHECKPOINT_NUMBER= Number from "model.ckpt-${CHECKPOINT_NUMBER}.meta"
python object_detection/export_inference_graph.py \
    --input_type=image_tensor \
    --pipeline_config_path=PATH_TO/LogosInTheWild-v2/models/model/faster_rcnn_inception_logos-locally-on-ubuntu.config \
    --trained_checkpoint_prefix=PATH_TO/LogosInTheWild-v2/models/model/model.ckpt-${CHECKPOINT_NUMBER} \
    --output_directory=PATH_TO/LogosInTheWild-v2/export

Open Logo Capstone Project/adjusted_object_detection_tutorial.ipynb in a jupyter notebook, adjust the paths and execute the cells.