sample_tracking_by_detection_and_landmark_localisation.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Load a video, apply a (face) detector and then track.\n",
    "\n",
    "# I do recommend either using the dlib detector if time \n",
    "# performance is crucial for you or the ffld2 (DPM model) \n",
    "# if you want the 'optimal' detection. A nice review of \n",
    "# why and how these options work for detection can be \n",
    "# found here: \n",
    "# http://ibug.doc.ic.ac.uk/media/uploads/documents/1603.06015v1.pdf "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from os.path import isdir, join\n",
    "import numpy as np\n",
    "\n",
    "import menpo.io as mio\n",
    "from menpodetect.ffld2 import load_ffld2_frontal_face_detector\n",
    "from menpodetect.dlib import load_dlib_frontal_face_detector\n",
    "from menpodetect.dlib.conversion import pointgraph_to_rect\n",
    "from dlib import shape_predictor\n",
    "from menpo.shape import PointCloud\n",
    "\n",
    "try:\n",
    "    %matplotlib inline\n",
    "    from menpowidgets import visualize_images\n",
    "except NameError:\n",
    "    # e.g. if you are in a terminal\n",
    "    pass"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Auxiliary functions for detector/landmark localisation method"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# decide which detector to use: \n",
    "# detector = load_ffld2_frontal_face_detector()\n",
    "detector = load_dlib_frontal_face_detector()\n",
    "\n",
    "# landmark localisation function\n",
    "def detection_to_pointgraph(detection):\n",
    "    return PointCloud(np.array([(p.y, p.x) for p in detection.parts()]))\n",
    "\n",
    "\n",
    "def landm_loc(im, model):\n",
    "    # accepts an image, a model (dlibERT) and performs \n",
    "    # the landmark localisation. The bounding box should exist.\n",
    "    im_pili = np.array(im.as_PILImage())\n",
    "    # the group is defined as the first label attached to the image.\n",
    "    group = im.landmarks.group_labels[0]\n",
    "    ln = im.landmarks[group]\n",
    "    det_frame = model(im_pili, pointgraph_to_rect(ln.lms))\n",
    "    init_pc = detection_to_pointgraph(det_frame)\n",
    "    im.landmarks['g'] = init_pc\n",
    "    return init_pc\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Base paths"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# base path of your clips\n",
    "pb = ('/vol/atlas/homes/grigoris/videos_external/'\n",
    "      'kollias_dimitris/2016_12_semaine/mp4/')\n",
    "assert isdir(pb), 'The base path does not exist.'\n",
    "# path to export the landmarks to\n",
    "p_out = 'path_to_export_the_landmarks'\n",
    "\n",
    "# base path of the landmark localisation model:\n",
    "# e.g. if you use the dlib detector, you can\n",
    "# download the following landmark localisation \n",
    "# model: \n",
    "# https://github.com/davisking/dlib-models\n",
    "p_sp = 'path_to_model'\n",
    "assert(isfile(p_sp))\n",
    "model = shape_predictor(p_sp)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Main processing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# import the video\n",
    "ims = mio.import_video(join(pb, 'Session3.2PrudenceOperator.mov'))\n",
    "# In case you have the frames, this should work by \n",
    "# replacing the 'import_video' with 'import_images' \n",
    "# in the appropriate paths."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# process all frames iteratively.\n",
    "for cnt, im in enumerate(ims):\n",
    "    if im.n_channels == 3:\n",
    "        im = im.as_greyscale()\n",
    "    lns = detector(im)\n",
    "\n",
    "    if im.landmarks.n_groups == 0:\n",
    "        # there are no detections\n",
    "        continue\n",
    "    name = '{0:06d}'.format(cnt)\n",
    "    im.constrain_landmarks_to_bounds()\n",
    "    # call the landmark localisation method.\n",
    "    landm_loc(im, model)\n",
    "    # export the landmarks\n",
    "    mio.export_landmark_file(im.landmarks['g'], \n",
    "                             p_out + name + '.pts', overwrite=True)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.4.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}