thesis-grizzly

Grizzly is a web-based application to search over a dataset of images based on two techniques:

• search by collage
• search by faces

To pull the repository:

git clone --depth 1 [email protected]:JankaSvK/thesis-grizzly.git

Firstly, download the demo data to play with. Replace current image_representations with the directory obtained from the zip.

The images used in the demo data are from Open Images Dataset. The images are under CC-by 4.0.

To run the application, Docker is recommended. Then run following commands:

$ docker build . -t app
$ docker run -p 8000:8000 -v $PWD/image_representations:/diplomova_praca/static/image_representations -t app

The application then can be accessed at: 127.0.0.1:8000

Obtaining data for custom dataset

(not necessary for running the demo)

It is possible to use the application over a custom dataset. To do that, firstly it is needed to obtain the features, which the application uses. It is possible to generate only some of them, all are not required. After obtaining the data, the original in images_representations can be replaced, or a new directory with the same structure can be created. In later case, change the path to the updated directory, when running the app.

Firstly, if the data are videos, please refer to a VIRET tool for extraction frames from videos. If they are images, with no temporal context, then you need to create one more subdirectory level.

images_representations/images/000/image1.jpg
images_representations/images/000/image2.jpg

If they have temporal context, you can split the data into corresponding directories. We do not rescale images for the frontend, they should be proprecessed to a resolution 320x180 (possible setting for VIRET tool).

Building docker for annotations

$ cd diplomova_praca_lib
$ docker build . -t lib

Regions Annotation for MobileNetV2

2x4 regions, PCA=128 components, input_size=96

$ images="/path/to/images/"
$ intermediate_output="/path/to/intermediate_output/"
$ features="/path/to/features/"

$ docker run \
  -v $images:/images \
  -v $intermediate_output:/feature_records \
  lib \
   python diplomova_praca_lib/annotate_images.py \
    --images_dir=/images --save_location=/feature_records \
    --feature_model=mobilenetv2 --num_regions=2,4 \
    --input_size=96

$ docker run \
  -v $intermediate_output:/feature_records \
  -v $features:/features \
  lib \
    python diplomova_praca_lib/preprocess_data.py \
      --input=/feature_records --output=/features 
      --transform --regions --explained_ratio=128

Faces Extractions

0.08 is the criteria on the size of face to be accepted.

$ images="/path/to/images/"
$ intermediate\_output="/path/to/intermediate_output/"
$ features="/path/to/features/"

$ docker run \
  -v $images:/images \
  -v $intermediate_output:/feature_records \
  lib \
   python diplomova_praca_lib/annotate_images.py \
    --images_dir=/images --save_location=/feature_records \
    --feature_model=faces

$ docker run \
  -v $intermediate\_output:/feature_records \
  -v $features:/features \
  lib \
    python diplomova_praca_lib/preprocess_face_data.py \
      --input=/feature_records --output=/features \
      --crop_size=0.08

Training SOM

$ face_features="/path/to/face_features/"
$ trained_som="/path/to/trained_som/"

$ docker run \
  -v $face_features:/features \
  -v $trained_som:/output \
  lib \
   python diplomova_praca_lib/train_som.py \
    --input=/features \
    --iterations=200000

Why grizzly?

Grizzlies spend most of their waking hours searching for food.