Code use to create COCO Attributes dataset and experiments in the associate ECCV 2016 paper.
Python JavaScript CSS HTML Shell
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
data first commit Oct 13, 2016
graph first commit Oct 13, 2016
.gitignore typos May 23, 2017
nginx.conf.bak first commit Oct 13, 2016

COCO Attributes annotation server and attribute classification code

Code use to create COCO Attributes dataset and experiments in the associated ECCV 2016 paper.

Before using this repository, create a in the top-level directory. Use as a guide.


Python requirements are listed in requirements.txt. To install use:

pip install -r requirements.txt

Additional requirements include:

  • Nginx (example config file in nginx.conf.bak)
  • Postgres
  • Caffe (for running classification experiements, not needed for annotation server)

COCO Attributes dataset

To download the current version of this dataset go to:

Instructions for starting the server

sudo service nginx start uwsgi --socket -w wsgi:app

Instructions for starting mturk hits ''' returns list of job_ids that have less than 3 hits by trusted workers job_id occur multiple times if missing multiple hits ''' get_missing_hits(job_type):

launch_missing_hits(job_type, task_file, mturk_rel_path) # Note: this module is called by, not by the requester directly in normal operation ''' this is the number of questions permitted in the ELA ''' NUMQ = 40 (set this variable in for max number of attributes to annotate per object) ''' func for adding (patch, attribute) question to Query table, check if enough queries to launch new hit for relaunching hits that already have labels ''' schedule_next_query(patch_id, cat_id) ''' Batches patch_ids into groups of 10 and attr_ids into groups of 20 Makes set of HITs allimgs=True for binary attribute labeling tasks with one image per query allimgs=False for multiple attribute annotation Output: list of job_ids created ''' make_tasks(patch_ids, attr_ids, task_label, job_type='annotation', allimgs=False) ''' server function that runs when a form from the 'allimgs' page is submitted these are assumed to be ela tasks, and the ela function is queried after the annotations are logged ''' allimgs_post(self, id)