single-rectangle.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:12.535000",
     "start_time": "2016-07-08T01:56:12.147000"
    },
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:14.210000",
     "start_time": "2016-07-08T01:56:13.200000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 8L, 8L), (50000L, 1L, 4L))"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Create images with random rectangles and bounding boxes. \n",
    "num_imgs = 50000\n",
    "\n",
    "img_size = 8\n",
    "min_object_size = 1\n",
    "max_object_size = 4\n",
    "num_objects = 1\n",
    "\n",
    "bboxes = np.zeros((num_imgs, num_objects, 4))\n",
    "imgs = np.zeros((num_imgs, img_size, img_size))  # set background to 0\n",
    "\n",
    "for i_img in range(num_imgs):\n",
    "    for i_object in range(num_objects):\n",
    "        w, h = np.random.randint(min_object_size, max_object_size, size=2)\n",
    "        x = np.random.randint(0, img_size - w)\n",
    "        y = np.random.randint(0, img_size - h)\n",
    "        imgs[i_img, x:x+w, y:y+h] = 1.  # set rectangle to 1\n",
    "        bboxes[i_img, i_object] = [x, y, w, h]\n",
    "        \n",
    "imgs.shape, bboxes.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:15.125000",
     "start_time": "2016-07-08T01:56:14.870000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAPsAAAEACAYAAAB1UADIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x75076d8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "i = 0\n",
    "plt.imshow(imgs[i].T, cmap='Greys', interpolation='none', origin='lower', extent=[0, img_size, 0, img_size])\n",
    "for bbox in bboxes[i]:\n",
    "    plt.gca().add_patch(matplotlib.patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3], ec='r', fc='none'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:15.930000",
     "start_time": "2016-07-08T01:56:15.803000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 64L), -6.8856031987252206e-17, 1.0000000000000009)"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Reshape and normalize the image data to mean 0 and std 1. \n",
    "X = (imgs.reshape(num_imgs, -1) - np.mean(imgs)) / np.std(imgs)\n",
    "X.shape, np.mean(X), np.std(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:16.558000",
     "start_time": "2016-07-08T01:56:16.548000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((50000L, 4L), 0.28105437500000002, 0.17477239890314314)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Normalize x, y, w, h by img_size, so that all values are between 0 and 1.\n",
    "# Important: Do not shift to negative values (e.g. by setting to mean 0), because the IOU calculation needs positive w and h.\n",
    "y = bboxes.reshape(num_imgs, -1) / img_size\n",
    "y.shape, np.mean(y), np.std(y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:17.198000",
     "start_time": "2016-07-08T01:56:17.192000"
    },
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Split training and test.\n",
    "i = int(0.8 * num_imgs)\n",
    "train_X = X[:i]\n",
    "test_X = X[i:]\n",
    "train_y = y[:i]\n",
    "test_y = y[i:]\n",
    "test_imgs = imgs[i:]\n",
    "test_bboxes = bboxes[i:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:56:21.616000",
     "start_time": "2016-07-08T01:56:17.964000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using Theano backend.\n"
     ]
    }
   ],
   "source": [
    "# Build the model.\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Activation, Dropout\n",
    "from keras.optimizers import SGD\n",
    "model = Sequential([\n",
    "        Dense(200, input_dim=X.shape[-1]), \n",
    "        Activation('relu'), \n",
    "        Dropout(0.2), \n",
    "        Dense(y.shape[-1])\n",
    "    ])\n",
    "model.compile('adadelta', 'mse')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:57:19.973000",
     "start_time": "2016-07-08T01:56:21.618000"
    },
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 40000 samples, validate on 10000 samples\n",
      "Epoch 1/30\n",
      "1s - loss: 0.0666 - val_loss: 0.0022\n",
      "Epoch 2/30\n",
      "1s - loss: 0.0047 - val_loss: 5.8258e-04\n",
      "Epoch 3/30\n",
      "1s - loss: 0.0019 - val_loss: 3.0037e-04\n",
      "Epoch 4/30\n",
      "1s - loss: 0.0014 - val_loss: 2.2738e-04\n",
      "Epoch 5/30\n",
      "1s - loss: 0.0012 - val_loss: 1.4762e-04\n",
      "Epoch 6/30\n",
      "2s - loss: 0.0011 - val_loss: 1.3302e-04\n",
      "Epoch 7/30\n",
      "2s - loss: 9.9042e-04 - val_loss: 1.1030e-04\n",
      "Epoch 8/30\n",
      "1s - loss: 9.1788e-04 - val_loss: 9.8579e-05\n",
      "Epoch 9/30\n",
      "1s - loss: 8.8779e-04 - val_loss: 8.4835e-05\n",
      "Epoch 10/30\n",
      "1s - loss: 8.5834e-04 - val_loss: 7.8099e-05\n",
      "Epoch 11/30\n",
      "1s - loss: 8.3532e-04 - val_loss: 7.2461e-05\n",
      "Epoch 12/30\n",
      "1s - loss: 8.0197e-04 - val_loss: 7.8811e-05\n",
      "Epoch 13/30\n",
      "2s - loss: 7.8899e-04 - val_loss: 7.1290e-05\n",
      "Epoch 14/30\n",
      "2s - loss: 7.6972e-04 - val_loss: 7.0324e-05\n",
      "Epoch 15/30\n",
      "1s - loss: 7.5325e-04 - val_loss: 5.9552e-05\n",
      "Epoch 16/30\n",
      "1s - loss: 7.4060e-04 - val_loss: 5.8210e-05\n",
      "Epoch 17/30\n",
      "1s - loss: 7.2943e-04 - val_loss: 6.9643e-05\n",
      "Epoch 18/30\n",
      "1s - loss: 7.1842e-04 - val_loss: 5.8081e-05\n",
      "Epoch 19/30\n",
      "1s - loss: 7.1059e-04 - val_loss: 5.3423e-05\n",
      "Epoch 20/30\n",
      "1s - loss: 6.9717e-04 - val_loss: 4.7243e-05\n",
      "Epoch 21/30\n",
      "2s - loss: 6.9394e-04 - val_loss: 6.2623e-05\n",
      "Epoch 22/30\n",
      "2s - loss: 6.8506e-04 - val_loss: 5.4235e-05\n",
      "Epoch 23/30\n",
      "2s - loss: 6.8581e-04 - val_loss: 4.9188e-05\n",
      "Epoch 24/30\n",
      "1s - loss: 6.7876e-04 - val_loss: 4.8977e-05\n",
      "Epoch 25/30\n",
      "1s - loss: 6.6712e-04 - val_loss: 4.3991e-05\n",
      "Epoch 26/30\n",
      "1s - loss: 6.6541e-04 - val_loss: 3.9585e-05\n",
      "Epoch 27/30\n",
      "1s - loss: 6.6264e-04 - val_loss: 4.2685e-05\n",
      "Epoch 28/30\n",
      "1s - loss: 6.4944e-04 - val_loss: 4.8632e-05\n",
      "Epoch 29/30\n",
      "1s - loss: 6.4864e-04 - val_loss: 5.2251e-05\n",
      "Epoch 30/30\n",
      "1s - loss: 6.4314e-04 - val_loss: 5.0742e-05\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x11195a20>"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Train.\n",
    "model.fit(train_X, train_y, nb_epoch=30, validation_data=(test_X, test_y), verbose=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:57:20.286000",
     "start_time": "2016-07-08T01:57:19.975000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000L, 1L, 4L)"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Predict bounding boxes on the test images.\n",
    "pred_y = model.predict(test_X)\n",
    "pred_bboxes = pred_y * img_size\n",
    "pred_bboxes = pred_bboxes.reshape(len(pred_bboxes), num_objects, -1)\n",
    "pred_bboxes.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T01:57:20.311000",
     "start_time": "2016-07-08T01:57:20.289000"
    },
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def IOU(bbox1, bbox2):\n",
    "    '''Calculate overlap between two bounding boxes [x, y, w, h] as the area of intersection over the area of unity'''\n",
    "    x1, y1, w1, h1 = bbox1[0], bbox1[1], bbox1[2], bbox1[3]\n",
    "    x2, y2, w2, h2 = bbox2[0], bbox2[1], bbox2[2], bbox2[3]\n",
    "\n",
    "    w_I = min(x1 + w1, x2 + w2) - max(x1, x2)\n",
    "    h_I = min(y1 + h1, y2 + h2) - max(y1, y2)\n",
    "    if w_I <= 0 or h_I <= 0:  # no overlap\n",
    "        return 0.\n",
    "    I = w_I * h_I\n",
    "\n",
    "    U = w1 * h1 + w2 * h2 - I\n",
    "\n",
    "    return I / U"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:01:30.216000",
     "start_time": "2016-07-08T02:01:29.481000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAArkAAACyCAYAAABGBtdeAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nrkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x12d42400>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Show a few images and predicted bounding boxes from the test dataset. \n",
    "plt.figure(figsize=(12, 3))\n",
    "for i_subplot in range(1, 5):\n",
    "    plt.subplot(1, 4, i_subplot)\n",
    "    i = np.random.randint(len(test_imgs))\n",
    "    plt.imshow(test_imgs[i].T, cmap='Greys', interpolation='none', origin='lower', extent=[0, img_size, 0, img_size])\n",
    "    for pred_bbox, exp_bbox in zip(pred_bboxes[i], test_bboxes[i]):\n",
    "        plt.gca().add_patch(matplotlib.patches.Rectangle((pred_bbox[0], pred_bbox[1]), pred_bbox[2], pred_bbox[3], ec='r', fc='none'))\n",
    "        plt.annotate('IOU: {:.2f}'.format(IOU(pred_bbox, exp_bbox)), (pred_bbox[0], pred_bbox[1]+pred_bbox[3]+0.2), color='r')\n",
    "        \n",
    "# plt.savefig('plots/bw-single-rectangle_prediction.png', dpi=300)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2016-07-08T02:01:45.826000",
     "start_time": "2016-07-08T02:01:45.738000"
    },
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.88166504246337463"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Calculate the mean IOU (overlap) between the predicted and expected bounding boxes on the test dataset. \n",
    "summed_IOU = 0.\n",
    "for pred_bbox, test_bbox in zip(pred_bboxes.reshape(-1, 4), test_bboxes.reshape(-1, 4)):\n",
    "    summed_IOU += IOU(pred_bbox, test_bbox)\n",
    "mean_IOU = summed_IOU / len(pred_bboxes)\n",
    "mean_IOU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}