frank-afn / clothing_segmentation

Tornado API to segment clothing in fashion images

Tornado API - Clothing Segmentation with opencv3

Tornado API to segment clothing in fashion images

Install

• Install the latest Anaconda distribtution to your system here (https://www.continuum.io/downloads)
  • Make sure you select the correct operating system
• Install OpenCv3 - There are good online walkthroughs. See Below
  • OSX - http://www.pyimagesearch.com/2015/06/15/install-opencv-3-0-and-python-2-7-on-osx/
  • OSX - https://www.learnopencv.com/install-opencv-3-on-yosemite-osx-10-10-x/
  • Ubuntu - http://www.pyimagesearch.com/2015/06/22/install-opencv-3-0-and-python-2-7-on-ubuntu/
• Clone this repository and create the app environment
  • $ mkdir cs_repo && cd cs_repo
  • $ git clone https://github.com/ncaadam/clothing_segmentation
  • $ cd clothing_segementation
  • $ CS_HOME=$(pwd)
  • $ conda create -p $CS_HOME/env -y --copy anaconda
• Make input and output folders in the repo folder ($CS_HOME)
  • $ mkdir input && mkdir output

Dataset

• Download the test dataset here -> (TBD)
• Extract the images to the $CS_HOME/input folder created above

Usage

• Navigate to the application, activate the environment, and start the API
  • $ cd $CS_HOME
  • $ source activate env
  • $ nohup python clothing_segmentation_api.py & (the API will run at localhost:9999)
• Send a POST request to the cut directory
  • $ curl -v -XPOST -H 'Content-Type: application/json' -d '{"num_threads": 8}' http://localhost:9999/cut

Findings

• As my laptop has 4 logical cores, the API runs most efficiently at 8 or 16 threads
  • 1 thread = 873ms/image
  • 2 threads = 453ms/image
  • 4 threads = 354ms/image
  • 8 threads = 327ms/image
  • 16 threads = 315ms/image
  • 32 threads = 330ms/image
  • 64 threads = 332ms/image

Future Considerations / Enhancements

• Pose estimation
• Scale invariant features
• Machine learning to identify the person
• More robust background and skin subtraction
• Add logging to a log file

Resources and Inspirations

• Automatic Segmentation of Clothing for the Identification of Fashion Trends Using K-Means Clustering - http://cs229.stanford.edu/proj2009/McDanielsWorsley.pdf
• Getting the Look: Clothing Recognition and Segmentation for Automatic Product Suggestions in Everyday Photos - http://image.ntua.gr/iva/files/kalantidis_icmr13.pdf
• 2D Articulated Human Pose Estimation and Retrieval in (Almost) Unconstrained Still Images - https://www.robots.ox.ac.uk/~vgg/publications/2012/Eichner12/eichner12.pdf