ready to merge

VoIP.py

@@ -1,29 +1,25 @@
 import cv2
-import urllib2
+import urllib.request as urllib2
 import numpy as np
 import sys
 
 # Input the URL shown by IP Webcam
-url = "35.3.102.178:8080"
-stream = 'http://' + url + '/video'
-print('Streaming from: ' + stream)
+url = "35.3.71.126:8080"
+tream = 'http://' + url + '/video'
+print('Streaming from: ' + tream)
 
 # Open ByteStram
-videoStream = urllib2.urlopen(stream)
+stream = urllib2.urlopen(tream)
 
-# Collect Bytes and Process
-packets = ''
+bytes = bytes()
 while True:
-    packets += videoStream.read(1024)
-    a = packets.find('\xff\xd8')
-    b = packets.find('\xff\xd9')
-    if a!=-1 and b!=-1:
-        jpg = packets[a:b+2]
-        packets = packets[b+2:]
-
-        # decode byte stream to form video
-        video = cv2.imdecode(np.fromstring(jpg, dtype=np.uint8),cv2.IMREAD_COLOR)
-        # display video
-        cv2.imshow('WebStream',video)
-        if cv2.waitKey(1) & 0xFF == ord('q'):
-            break
+    bytes += stream.read(1024)
+    a = bytes.find(b'\xff\xd8')
+    b = bytes.find(b'\xff\xd9')
+    if a != -1 and b != -1:
+        jpg = bytes[a:b+2]
+        bytes = bytes[b+2:]
+        i = cv2.imdecode(np.fromstring(jpg, dtype=np.uint8), cv2.IMREAD_COLOR)
+        cv2.imshow('i', i)
+        if cv2.waitKey(1) == 27:
+            exit(0)

data_handler.py

@@ -21,7 +21,7 @@ class Process:
     remsamples() = prints total number of samples that remain from the selected lot of images
     '''
 
-    def __init__(self, location, height = 224, width = 224, depth = 3):
+    def __init__(self, location, file = None, flag = False, height = 224, width = 224, depth = 3):
         self.img2labels = dict()
         self.idx2img = dict()
         self.num_crops=1
@@ -37,17 +37,24 @@ def __init__(self, location, height = 224, width = 224, depth = 3):
         self.remImages = []
         self.remSamples = []
 
+        # flag for processing data from NCLT / Cambridge
+        self.flag = flag
+        self.file = file
         # OS Walk for getting all image files
         self.OSWalk()
         # print number of images
         self.numimages()
-        # populate labels dict
-        self.getLabels()
 
+        # populate labels dict
+        # False implies Cambridge Dataset, True is NCLT
+        if self.flag == False:
+            self.getLabels()
+        else:
+            self.getGround()
 
     def OSWalk(self):
         images = [os.path.join(root, name) for root, dirs, files in os.walk(self.location)
-             for name in files if name.endswith((".png", ".jpg", ".jpeg", ".gif"))]
+             for name in files if name.endswith((".png", ".jpg", ".jpeg", ".gif", ".tiff"))]
         self.imageLocs = images
 
         # for i in range(len(images)):
@@ -58,8 +65,6 @@ def OSWalk(self):
         self.numImages = len(self.imageLocs)
         self.remImages = np.ones((self.numImages), dtype = bool)
 
-
-
     def generateData(self, num):
         ctr = 0
         indices = ['']*num
@@ -87,22 +92,22 @@ def generateData(self, num):
         self.process(num, indices)
 
     def centeredCrop(self, img, output_side_length):
-	height, width, depth = img.shape
-	new_height = output_side_length
-	new_width = output_side_length
-	if height > width:
-		new_height = output_side_length * height / width
-	else:
-		new_width = output_side_length * width / height
-	height_offset = (new_height - output_side_length) / 2
-	width_offset = (new_width - output_side_length) / 2
-	cropped_img = img[height_offset:height_offset + output_side_length,
-						width_offset:width_offset + output_side_length]
-	return cropped_img
+        height, width, depth = img.shape
+        new_height = output_side_length
+        new_width = output_side_length
+        if height > width:
+            new_height = output_side_length * height / width
+        else:
+            new_width = output_side_length * width / height
+            height_offset = (new_height - output_side_length) / 2
+            width_offset = (new_width - output_side_length) / 2
+            cropped_img = img[height_offset:height_offset + output_side_length,
+                              width_offset:width_offset + output_side_length]
+        return cropped_img
 
 
     def process(self, num, indices):
-        print('Generating Samples .... ')
+        print('Generating Samples .... Note: no img normalization ')
 
         self.sampleImages = np.zeros((num*self.num_crops , self.height, self.width, self.depth), dtype=np.uint8)
 
@@ -117,6 +122,7 @@ def process(self, num, indices):
             names.append(name)
         means = np.mean(imgs, axis=0)
         imgs = imgs - means
+
         #temp_mean = np.mean(img, axis=0)
         #temp_mean = np.mean(temp_mean, axis=0)
         #temp_std = np.zeros(self.depth)
@@ -127,12 +133,19 @@ def process(self, num, indices):
         #    img[:,:,i] /= temp_std[i]
 
         for i in range(num):
+            # rotate image if NCLT
+            img = imgs[i, :,:,:]
+            if self.flag==True:
+                rows, cols, _ = img.shape
+                M = cv2.getRotationMatrix2D((cols/2,rows/2),90,1)
+                img = cv2.warpAffine(img,M,(cols,rows))
+
             # generate 128 random indices for crop
             for j in range(idx, idx+self.num_crops ):
                 x = randint(0,31)
                 y = randint(0,230)
-                img = imgs[i, :,:,:]
-                self.sampleImages[j, :, :, :] = self.centeredCrop(img, 224)#img[x:x + self.height, y:y + self.width, :].copy()
+
+                self.sampleImages[j, :, :, :] = img[x:x + self.height, y:y + self.width, :].copy()
                 self.idx2img[j] = names[i]
             idx += self.num_crops
 
@@ -157,11 +170,13 @@ def fetch(self, num):
                     self.remSamples[idx] = False
 
                     # gaussian normalization of image to have mean 0, variance 1
-                    temp = self.sampleImages[idx, :, :, :].astype(float)
                     try:
                         labels[ctr] = self.img2labels[self.idx2img[idx]]
+                        samples[ctr,:,:,:] = self.sampleImages[idx, :, :, :].astype(float)
+
                     except KeyError:
                         pdb.set_trace()
+                        continue
                     ctr += 1
 
         return [flag, samples, labels]
@@ -183,15 +198,39 @@ def getLabels(self):
             line = lines[i].strip()
             line = line.split()
             self.img2labels[line[0]] = list(map(float,line[1:8]))
-
-    def getName(self, loc):
-        ctr = 0
-        for i in range(len(loc)):
-            if loc[i]=='/':
-                ctr += 1
-            if ctr==2:
-                return loc[i+1:]
 
+    def getGround(self):
+        filename = self.location + self.file
+        odom = np.loadtxt(filename, delimiter = ",")
+        length, _ = odom.shape
+
+        # the number we have to divide by to get a correct association
+        param = 4
+
+        for i in range(length):
+            name = int(odom[i, 0].item()/10**param)
+
+            x = float(odom[i, 1])
+            y = float(odom[i, 2])
+            z = float(odom[i, 3])
+            r = float(odom[i, 4])
+            p = float(odom[i, 5])
+            h = float(odom[i, 6])
+            data = [x,y,z,r,p,h]
+            self.img2labels[name] = data
+
+    def getName(self, loc):
+        if self.flag == False:
+            ctr = 0
+            for i in range(len(loc)):
+                if loc[i]=='/':
+                    ctr += 1
+                if ctr==2:
+                    return loc[i+1:]
+        else:
+            params = loc.split('/')
+            name = params[-1][:-9]
+            return int(name)
 
 
     def reset(self):
@@ -217,4 +256,4 @@ def remsamples(self):
 
     def remimages(self):
         print('The number of images that remain are: ',np.sum(self.remImages))
-        return remImages
+        return self.remImages

test.py

@@ -1,7 +1,6 @@
 import data_handler as DH
-import sys
 import numpy as np
-
+import cv2
 '''
 Usage:
 location : enter location of main directory
@@ -13,25 +12,34 @@
 remimages() = prints total number of images that remain in the directory
 remsamples() = prints total number of samples that remain from the selected lot of images
 '''
-print(sys.version)
 
-location = "../ShopFacade/"
+# location = "./2013-01-10/"
+# file = 'groundtruth_2013-01-10.csv'
+
+location = "/media/eecs568/My Passport/NCLT/2012-03-17/2012-03-17/"
+file = 'groundtruth_2012-03-17.csv'
 
-dh = DH.Process(location)
+dh = DH.Process(location, file, True)
 
 # pick number of images to pick samples from
-numImages = 50
+numImages = 100
 dh.generateData(numImages)
 dh.remimages()
 
 dh.numsamples()
-flag, images, labels = dh.fetch(60)
+flag, images, labels = dh.fetch(2)
+print(labels)
 print(images.shape, len(labels), flag)
 dh.remsamples()
 
+
 # Trick to pick samples from selected images
 numSamples = 60
 flag = True
 while flag==True:
     flag, images, labels = dh.fetch(numSamples)
+    cv2.imshow('image', images[0,:,:,:])
+    cv2.waitKey(0)
+    pdb.set_trace()
     dh.remsamples()
+

train.py

@@ -29,14 +29,15 @@ def delete_network_backups(filename_prefix):
 
 class trainer():
 
-    def __init__(self,path_to_weight, path_to_data, beta, resume_training=False):
+    def __init__(self,path_to_weight, path_to_data, beta, use_quaternion=True, resume_training=False):
         self.network_input_size = 224
+        self.output_dim = 7 if use_quaternion else 6
         self.sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
         if resume_training:
             self.restore_network(path_to_weight)
         else:
             self.image_inputs = tf.placeholder(tf.float32, [None, self.network_input_size, self.network_input_size, 3])
-            self.label_inputs = tf.placeholder(tf.float32, [None, 7])  # [ X Y Z W  P Q R]
+            self.label_inputs = tf.placeholder(tf.float32, [None, self.output_dim])  # [ X Y Z W  P Q R]
 
             self.network = vgg.VGG16({'data': self.image_inputs})
             self.regen_regression_network()
@@ -84,9 +85,9 @@ def regen_regression_network(self):
             fc9_input_shape = fc8.get_shape()
 
             feed_in, dim = (fc8, fc9_input_shape[-1].value)
-            fc9_weights = tf.get_variable('weights', [dim, 7])
+            fc9_weights = tf.get_variable('weights', [dim, self.output_dim])
             self.network.variable_summaries( fc9_weights,  "_weights_fc" )
-            fc9_biases =  tf.get_variable('biases', [7])
+            fc9_biases =  tf.get_variable('biases', [self.output_dim])
             op = tf.nn.xw_plus_b
             fc9 = op(feed_in, fc9_weights, fc9_biases, name=scope.name)
         self.init_vars = [fc8_weights, fc8_biases , fc9_weights,  fc9_biases]
@@ -109,8 +110,7 @@ def restore_network(self, path_to_weight):
 
     def build_loss(self, beta=100):
         self.translation_loss = tf.sqrt(tf.nn.l2_loss(self.regression_out[0:3] - self.label_inputs[0:3]))
-        self.rotation_loss = tf.sqrt(tf.nn.l2_loss( self.regression_out[3:7] - self.label_inputs[3:7]  ))
-        #self.rotation_loss = tf.sqrt(tf.nn.l2_loss( self.regression_out[3:7]/tf.norm(self.regression_out[3:7]) - self.label_inputs[3:7]  ))
+        self.rotation_loss = tf.sqrt(tf.nn.l2_loss( self.regression_out[3:] - self.label_inputs[3:]  ))
         self.loss = self.translation_loss + beta * self.rotation_loss
         tf.identity(self.loss, name="final_loss")