TOC

• Setup the environment
• Services
  • Start services
  • URLs to access the services
• Generate online dense vector model

TITAN

Typing-free Image and Tag Augmented Navigation

Setup the environment

Miniconda

We use Miniconda to handle the compatibility between the python packages and make the virtual pyhton env.

• download
• install
• getting started with conda

Install packages

conda install django==1.11.10
conda install -c conda-forge djangorestframework
conda install keras
conda install numpy
conda install opencv
conda install scikit-image
conda install elasticsearch==5.4
conda install elasticsearch-dsl==5.3

Install keras-gpu

• able to run keras on GPU

  conda install keras-gpu

• verify

  import keras
  keras.backend.tensorflow_backend._get_available_gpus()

Services

Start services

There are 2 services need to start:

• Django - the web server for image search
• Elasticsearch - the search engine for tag search

# Image search
sudo su -l titan 
cd TITAN/image_search/fileupload/
nohup python manage.py runserver 0.0.0.0:8000 &

# tag navigation
sudo su -l elk
cd elk/elasticsearch/
./bin/elasticsearch -d

URLs to access the services

Demo site

http://venraas.github.io/demo/recomd_mobile_20190408_package/recomd_tags_cam.html

image search simple demo

http://${ServerIP}:8000/cupid/upload/

elasticsearch (chrome app - Elasticsearch Head)

http://${ServerIP}:9200

Generate Online Dense Vector Model

In this step, we generate the images dense vectors, flat_imgfea.py, and concatenate them, compact_feature_dir.py, into multiple compact bulks with numpy format.

We will apply this dense vector model to calculate the similarity with the input image and seek the nearest topK for result.

flat_imgfea.py

Generates the model dense vectors for all input images.
Ex. python flat_imgfea.py n04204238-shopping_basket vgg16_dense

usage: flat_imgfea.py [-h] dirImgs dirFeaVcts

positional arguments:
  dirImgs     the directory of the source images
  dirFeaVcts  the directory for the output feature vectors of the given images

optional arguments:
  -h, --help  show this help message and exit

compact_feature_dir.py

Concatenates a bunch of the image feature vectors into a compact bulk vector.
Ex. python compact_feature_dir.py vgg16_dense

usage: compact_feature_dir.py [-h] dir

positional arguments:
  dir         the root dir of all image features

optional arguments:
  -h, --help  show this help message and exit