Convert MS COCO Annotation to Pascal VOC format

coco.m

function coco(n, k, note, dotrainval, testyear)
% Train and evaluate a model. 
%   pascal(n, note, dotrainval, testyear)
%
%   The model will be a mixture of n star models, each of which
%   has 2 latent orientations.
%
% Arguments
%   n             Number of aspect ratio clusters to use
%                 (The final model has 2*n components)
%   note          Save a note in the model.note field that describes this model
%   dotrainval    Also evaluate on the trainval dataset
%                 This is used to collect training data for context rescoring
%   testyear      Test set year (e.g., '2007', '2011')

% AUTORIGHTS
% -------------------------------------------------------
% Copyright (C) 2011-2012 Ross Girshick
% Copyright (C) 2008, 2009, 2010 Pedro Felzenszwalb, Ross Girshick
% 
% This file is part of the voc-releaseX code
% (http://people.cs.uchicago.edu/~rbg/latent/)
% and is available under the terms of an MIT-like license
% provided in COPYING. Please retain this notice and
% COPYING if you use this file (or a portion of it) in
% your project.
% -------------------------------------------------------

startup;

conf = voc_config();
cachedir = conf.paths.model_dir;
testset = conf.eval.test_set;
load('categories.mat')

% TODO: should save entire code used for this run
% Take the code, zip it into an archive named by date
% print the name of the code archive to the log file
% add the code name to the training note
timestamp = datestr(datevec(now()), 'dd.mmm.yyyy:HH.MM.SS');

% Set the note to the training time if none is given
if nargin < 3
  note = timestamp;
end

% Don't evaluate trainval by default
if nargin < 4
  dotrainval = false;
end

if nargin < 5
  % which year to test on -- a string, e.g., '2007'.
  testyear = conf.pascal.year;
end


for i = k:8:80
  % Record a log of the training and test procedure
  cls = categories(i, :)
  diary(conf.training.log([cls '-' timestamp]));

  % Train a model (and record how long it took)
  th = tic;
  model = pascal_train(cls, n, note);
  toc(th);

  % Free the feature vector cache memory
  fv_cache('free');

  % Lower threshold to get high recall
  model.thresh = min(conf.eval.max_thresh, model.thresh);
  model.interval = conf.eval.interval;

  suffix = testyear;

  % Collect detections on the test set
  ds = pascal_test(model, testset, testyear, suffix);

  % Evaluate the model without bounding box prediction
  ap1 = pascal_eval(cls, ds, testset, testyear, suffix);
  fprintf('AP = %.4f (without bounding box prediction)\n', ap1)

  % Recompute AP after applying bounding box prediction
  [ap1, ap2] = bboxpred_rescore(cls, testset, testyear, suffix);
  fprintf('AP = %.4f (without bounding box prediction)\n', ap1)
  fprintf('AP = %.4f (with bounding box prediction)\n', ap2)

  % Compute detections on the trainval dataset (used for context rescoring)
  if dotrainval
    trainval(cls);
  end
end

coco2pascal.py

import baker
import json
from path import path
from cytoolz import merge, join, groupby
from cytoolz.compatibility import iteritems
from cytoolz.curried import update_in
from itertools import starmap
from collections import deque
from lxml import etree, objectify
from scipy.io import savemat
from scipy.ndimage import imread


def keyjoin(leftkey, leftseq, rightkey, rightseq):
    return starmap(merge, join(leftkey, leftseq, rightkey, rightseq))


def root(folder, filename, height, width):
    E = objectify.ElementMaker(annotate=False)
    return E.annotation(
            E.folder(folder),
            E.filename(filename),
            E.source(
                E.database('MS COCO 2014'),
                E.annotation('MS COCO 2014'),
                E.image('Flickr'),
                ),
            E.size(
                E.width(width),
                E.height(height),
                E.depth(3),
                ),
            E.segmented(0)
            )


def instance_to_xml(anno):
    E = objectify.ElementMaker(annotate=False)
    xmin, ymin, width, height = anno['bbox']
    xmin = int(xmin)
    ymin = int(ymin)
    width = int(width)
    height = int(height)

    return E.object(
            E.name(anno['category_id']),
            E.bndbox(
                E.xmin(xmin),
                E.ymin(ymin),
                E.xmax(xmin+width),
                E.ymax(ymin+height),
                ),
            )


@baker.command
def write_categories(coco_annotation, dst):
    content = json.loads(path(coco_annotation).expand().text())
    categories = tuple( d['name'] for d in content['categories'])
    savemat(path(dst).expand(), {'categories': categories})


def get_instances(coco_annotation):
    coco_annotation = path(coco_annotation).expand()
    content = json.loads(coco_annotation.text())
    categories = {d['id']: d['name'] for d in content['categories']}
    return categories, tuple(keyjoin('id', content['images'], 'image_id', content['annotations']))

def rename(name, year=2014):
        out_name = path(name).stripext()
        # out_name = out_name.split('_')[-1]
        # out_name = '{}_{}'.format(year, out_name)
        return out_name


@baker.command
def create_imageset(annotations, dst):
    annotations = path(annotations).expand()
    dst = path(dst).expand()
    val_txt = dst / 'val.txt'
    train_txt = dst / 'train.txt'

    for val in annotations.listdir('*val*'):
        val_txt.write_text('{}\n'.format(val.basename().stripext()), append=True)

    for train in annotations.listdir('*train*'):
        train_txt.write_text('{}\n'.format(train.basename().stripext()), append=True)

@baker.command
def create_annotations(dbpath, subset, dst):
    annotations_path = path(dbpath).expand() / 'annotations/instances_{}2014.json'.format(subset)
    images_path = path(dbpath).expand() / 'images/{}2014'.format(subset)
    categories , instances= get_instances(annotations_path)
    dst = path(dst).expand()

    for i, instance in enumerate(instances):
        instances[i]['category_id'] = categories[instance['category_id']]

    for name, group in iteritems(groupby('file_name', instances)):
        img = imread(images_path / name)
        if img.ndim == 3:
            out_name = rename(name)
            annotation = root('VOC2014', '{}.jpg'.format(out_name), 
                              group[0]['height'], group[0]['width'])
            for instance in group:
                annotation.append(instance_to_xml(instance))
            etree.ElementTree(annotation).write(dst / '{}.xml'.format(out_name))
            print out_name
        else:
            print instance['file_name']





if __name__ == '__main__':
    baker.run()