Merge branch 'master' of https://github.com/Microsoft/arcticseals

src/AccuracyAssessment/validateThermal.py

@@ -25,6 +25,7 @@
 falseNegatives = 0
 hotSpotDistances = []
 registrationDistances = []
+truePositives = 0
 classificationTrue = 0
 classificationFalse = 0
 expectedTruePositives = 0
@@ -77,7 +78,7 @@ def getSortKey(item):
         for key in resDict:
             # if photo not in expected dictionary that means there were no animals in the entire photo so throw FalsePositive
             if key not in expectedDict:
-                falsePositives+=1
+                falsePositives+=len(resDict[key])
             else:
                 resRows = resDict.get(key)
                 expectedRows = expectedDict.get(key)
@@ -107,50 +108,53 @@ def getSortKey(item):
                 if (len(expectedRowsData) > 1):
                     expectedRowsData.sort(key=getSortKey)
 
-                for m in range(minLength):
-                    resRow = resRowsData[m]
-                    expectedRow = expectedRowsData[m]
-
-                    # 1 is expected values; 2 is ouput values. Both are location on the image
-                    x1 = int(expectedRow[5])
-                    y1 = int(expectedRow[6])
-                    x2 = int(resRow[5])
-                    y2 = int(resRow[6])
-                    dist = math.hypot(x2 - x1, y2 - y1)
-
-                    hotSpotDistances.append(dist)
-                    # The average size of seal on thermal image is about 2 by 5 pixels. Assume hot spot detedted within 
-                    # the size of a seal is a success detection.
-                    if (dist < locationOffsetTolerance):
-                        hotspotLocationTrue+=1
+                for e in expectedRowsData:
+                    minDist = 99999999
+                    for r in resRowsData:
+                        if r[1] != 'MatchesGroundTruth':
+                            x1 = int(expectedRow[5])
+                            y1 = int(expectedRow[6])
+                            x2 = int(resRow[5])
+                            y2 = int(resRow[6])
+                            dist = math.hypot(x2 - x1, y2 - y1)
+                            if dist < minDist:
+                                minDist = dist
+                                minResRow = r
 
-                    # classification accuracy
-                    if (expectedRow[12] == resRow[12]):
-                        classificationTrue += 1 
+                    # Was there a detected hotspot within tolerance of this ground truth hotspot?
+                    if minDist < locationOffsetTolerance:
+                        truePositives += 1
+                        minResRow[1] = 'MatchesGroundTruth' # Hackily reuse the "timestamp" field to indicate that we have a match
+                        if minResRow[13] == e[13]:
+                            classificationTrue += 1
+                        else
+                            classificationFalse += 1
+
+                for r in resRowsData:
+                    if r[1] != 'MatchesGroundTruth':
+                        falsePositives += 1
                     else:
-                        classificationFalse += 1
+                        if r[13] == 
+
 
     # Print results on screen
     print("Accuracy assessment for thermal image classification")
     print("Number of hot spots found: " + str(len(resList)-1))
-    print("Percent of hot spots found: " + str((classificationTrue + classificationFalse) / expectedTruePositives*100)  + " percent.")
+    print("Percent of hot spots found: " + str(truePositives / expectedTruePositives*100)  + " percent.")
     print("There were " + str(falsePositives) + " false positives.")
     print("There were " + str(falseNegatives) + " false negatives.")
-    if ((classificationTrue + classificationFalse) > 0):
+    if ((truePositives) > 0):
         print("There was a classification accuracy of " + str((classificationTrue / (classificationTrue + classificationFalse)) * 100) + " percent.")
-    if (len(hotSpotDistances) > 0):
-        print("There was an average hot spot distance of " + str(sum(hotSpotDistances) / len(hotSpotDistances)))        
-        print("There was an hot spot location detection accuracy of " + str(hotspotLocationTrue / len(hotSpotDistances)*100)  + " percent.")
 
     # Print result to a file
     with open('assessmentResults.txt', 'w') as f:
         print("Accuracy assessment for thermal image classification", file=f)
         print("Number of hot spots found: " + str(len(resList)-1), file=f)
-        print("Percent of hot spots found: " + str((classificationTrue + classificationFalse) / expectedTruePositives*100)  + " percent.", file=f)
+        print("Percent of hot spots found: " + str((truePositives) / expectedTruePositives*100)  + " percent.", file=f)
         print("There were " + str(falsePositives) + " false positives.", file=f)
         print("There were " + str(falseNegatives) + " false negatives.", file=f)
-        if ((classificationTrue + classificationFalse) > 0):
-            print("There was a classification accuracy of " + str((classificationTrue / (classificationTrue + classificationFalse)) * 100) + " percent.", file=f)
+        if ((truePositives) > 0):
+            print("There was a classification accuracy of " + str((truePositives / (truePositives)) * 100) + " percent.", file=f)
         if (len(hotSpotDistances) > 0):
             print("There was an average hot spot distance of " + str(sum(hotSpotDistances) / len(hotSpotDistances)), file=f)        
             print("There was an hot spot location detection accuracy of " + str(hotspotLocationTrue / len(hotSpotDistances)*100)  + " percent.", file=f)

src/ir-hotspot-rfc/README.md

@@ -1,3 +1,43 @@
 ## Hotspot Classifier based on Random Forest Classifier
 
-Will be filled in in next commit
+This folder contains the code for training and applying a random forest classifier in sci-kit learn to the arcticseal data. The classifier uses Sci-Kit Learn's classifier class so they can easily be switched out to test other classifiers. We are also using PCA to project the raw data onto "eigen-seals" like the following table:
+|Eigen Seal 1 | Eigen Seal 2|
+|------------ | ------------|
+|![alt text](./images/eigenseal1.png "Eigen-Seal 1") | ![alt text](./images/eigenseal2.png "Eigen-Seal 2")|
+
+## Classifying
+The classification script supports two modes for processing images: command line and as a library import to another script.
+
+#### Launch Classifier from the Command line
+The script requires three flags from the command line with an optional fourth:
+1. `--datadir` the path to the directory where the thermal images are stored
+1. `--datafile` the path to the csv file that you want to process
+1. `--modelfile` the path to the pickled version of the classifier (output by hotspot_training.py)
+1. `--outfile` (optional) the path where the csv with classification information should be written. Defaults to `./output.csv`
+
+The classified CSV file will have the column `hotspot_type` filled in with either "Anamoly" or "Animal" depending on the classification. It will also fill in the column `ir_confidence` that contains the confidence that the classifier has in its decision.
+
+An example of running the script is shown below:
+```bash
+python hotspot_classifier.py --datadir ./ArcticSealsData01_Thermal_N/ --datafile ../arcticseals/data/test.csv --modelfile pca_rfc_model_20180725_154906.p
+```
+
+#### Use Classifier as a Import
+The main function for classifying is called `classify_data` and takes the same input values that are passed from the command line as arguments. The function definition and doc string are shown here for convenience:
+```python
+def classify_data(data_file, data_directory, model_file, output_file):
+    """Data loading, classifying and output logic. For compatibility with library inputs
+    Inputs:
+        data_file: string, path to the input csv file
+        data_directory: string, path to the thermal images
+        model_file: string, path to the classifier model
+        output_file: string, path to the output csv file
+    """
+```
+
+## File Description
+* hotspot_classifier.py: Function for applying the model to a new set of data
+* hotspot_training.py: Script for training the model
+* pca_rfc_model_20180725_154906.p: Pre-trained model on arcticseals/data/train.csv
+* requirements.txt: pip compatible file listing the required dependencies. Install using `pip install requirements.txt`
+* README.md: this file

src/viewer/app/app.js

@@ -6,6 +6,7 @@ angular.module('myApp', [
   'myApp.view1',
   'myApp.view2',
   'myApp.view3',  
+  'myApp.view4',  
   'myApp.version'
 ]).
 config(['$locationProvider', '$routeProvider', function($locationProvider, $routeProvider) {

src/viewer/app/index.html

@@ -6,7 +6,7 @@
 <head>
   <meta charset="utf-8">
   <meta http-equiv="X-UA-Compatible" content="IE=edge">
-  <title>My AngularJS App</title>
+  <title>Find Seal in Artic with Deep Learning</title>
   <meta name="description" content="">
   <meta name="viewport" content="width=device-width, initial-scale=1">
   <link rel="stylesheet" href="bower_components/html5-boilerplate/dist/css/normalize.css">
@@ -24,6 +24,7 @@
     <li><a href="#!/view1">Color</a></li>
     <li><a href="#!/view2">Thermal</a></li>
     <li><a href="#!/view3">Area of Interest</a></li>    
+    <li><a href="#!/view4">Aggregate</a></li>    
   </ul>
 
   <div ng-view></div>
@@ -34,6 +35,7 @@
   <script src="view1/view1.js"></script>
   <script src="view2/view2.js"></script>
   <script src="view3/view3.js"></script>
+  <script src="view4/view4.js"></script>
   <script src="components/version/version.js"></script>
   <script src="components/version/version-directive.js"></script>
   <script src="components/version/interpolate-filter.js"></script>

src/viewer/app/view3/view3.js

@@ -1,7 +1,7 @@
 'use strict';
 
 var app = angular.module('myApp.view3', ['ngRoute'])
-var artic;
+var artic, items;
 var getName = function(articI) {
 	var names = articI.filt_color.split(".");
 	var name = names[0] + "." + names[1] + "_HOTSPOT_" + articI.hotspot_id + ".JPG";
@@ -22,7 +22,7 @@ app.controller('View3Ctrl', function($scope) {
 	$.getJSON( "/data/training.json", function( data ) {
 		if (data.Artic !== 'undefined') {
 			artic = data.Artic;
-			var len = artic.length;
+			var len = artic.length / 3;
 			for(var i = 0; i < len; i++) {
 				var name = getName(artic[i]);
 				$scope.items.push({"idx": i, "file": "crop-img//" + name});
@@ -34,13 +34,14 @@ app.controller('View3Ctrl', function($scope) {
 	$scope.filterMe = function( hotspot_type ) {
 		if (artic !== 'undefined') {
 			$scope.items = [];
-			var len = artic.length;
+			var len = artic.length / 3;
 			for(var i = 0; i < len; i++) {
 				var name = getName(artic[i]);
 				if (artic[i].hotspot_type === hotspot_type) {
 					$scope.items.push({"idx": i, "file": "crop-img//" + name});				
 				}
 			}
+			items = $scope.items;
 			$scope.$apply();
 		}
 	};

src/viewer/app/view4/view4.html

@@ -0,0 +1,14 @@
+<button class="btn btn-secondary dropdown-toggle" type="button" id="dropdownMenuButton" data-toggle="dropdown" aria-haspopup="true" aria-expanded="false">
+	Filter:
+</button>
+<div class="dropdown-menu">
+  <a class="dropdown-item" ng-click="filterMe('Anomaly')">Anomaly</a>
+  <a class="dropdown-item" ng-click="filterMe('Animal')">Animal</a>
+</div>
+
+<div class="row">
+	<div ng-repeat="item in items">
+		<img src="{{item.color}}"></img>
+		<img src="{{item.thermal}}"></img><br/>
+	</div>
+</div>

src/viewer/app/view4/view4.js

@@ -0,0 +1,50 @@
+'use strict';
+
+var app = angular.module('myApp.view4', ['ngRoute'])
+var artic, items;
+var getName = function(articI) {
+	var names = articI.filt_color.split(".");
+	var name = names[0] + "." + names[1] + "_HOTSPOT_" + articI.hotspot_id + ".JPG";
+	return name;
+};
+
+app.config(['$routeProvider', function($routeProvider) {
+  $routeProvider.when('/view4', {
+    templateUrl: 'view4/view4.html',
+    controller: 'View4Ctrl'
+  });
+}])
+
+app.controller('View4Ctrl', function($scope) {
+	$scope.items = [];
+
+	var json = null;
+
+	$.getJSON( "/data/training.json", function( data ) {
+		if (data.Artic !== 'undefined') {
+			var artic = data.Artic;
+			var len = artic.length / 2;
+			for(var i = 0; i < len; i++) {
+				var articI = artic[i];
+				$scope.items.push({"idx": i, "color": "thumb-img//" + articI.filt_color, "thermal" : "thumb-thermal//" + articI.filt_thermal8});
+			}
+			$scope.$apply();
+		}
+	});
+
+	$scope.filterMe = function( hotspot_type ) {
+		if (artic !== 'undefined') {
+			$scope.items = [];
+			var len = artic.length / 3;
+			for(var i = 0; i < len; i++) {
+				var name = getName(artic[i]);
+				if (artic[i].hotspot_type === hotspot_type) {
+					$scope.items.push({"idx": i, "file": "crop-img//" + name});				
+				}
+			}
+			items = $scope.items;
+			$scope.$apply();
+		}
+	};
+
+});

src/viewer/server.js

@@ -3,6 +3,9 @@ const app = express();
 const router = express.Router();
 const $ = jQuery = require('./jquery.js');
 var csv = require('./jquery.csv.js');
+var exec = require('child_process').exec;
+const fs = require('fs');
+
 const port = 3000;
 
 app.use('/', express.static('app'));
@@ -19,9 +22,16 @@ app.use('/thumb-thermal', express.static('E:\\noaa\\thumb-thermal'));
 app.use('/crop-img', express.static('E:\\noaa\\crop-img'));
 
 // using router.get() to prefix our path
-// url: http://localhost:3000/api/
-app.get('/api', (request, response) => {
-  response.json({message: 'Hello, welcome to my server'});
+// url: http://localhost:3000/api/CHESS_FL1_C_160408_000830.513_COLOR-8-BIT_HOTSPOT_17486.JPG
+app.get('/api/:id', (request, response) => {
+	var filename = 'E:\\noaa\\crop-img\\' + request.params.id;
+	fs.copyFileSync(filename, 'D:\\github\\seals\\src\\viewer\\util\\cnn\\cur\\cur.jpg');
+	console.log(request.params.id);
+	exec('python D:\\github\\seals\\src\\viewer\\util\\cnn\\SealCurrent.py', {cwd:'D:\\github\\seals\\src\\viewer\\util\\cnn'}, function callback(error, stdout, stderr){
+		var lines = stdout.split('\n');
+		//console.log(lines[4]);
+		response.json({message: lines[4]});
+	});
 });
 
 // set the server to listen on port 3000

src/viewer/util/Convert CSV to JSON.ipynb

@@ -71,7 +71,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.5"
+   "version": "3.5.5"
   }
  },
  "nbformat": 4,

src/viewer/util/Crop Images.ipynb

@@ -81,7 +81,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.5"
+   "version": "3.5.5"
   }
  },
  "nbformat": 4,

src/viewer/util/conv16.py

@@ -11,6 +11,7 @@
 for currentFile in inputPath.glob("*.png"):
     currentFilename = str(currentFile.name)
     inputFile = os.path.join(inputDir, currentFilename)
-    img = cv2.imread(inputFile, -cv2.IMREAD_ANYDEPTH).astype(np.uint8)
+    #img = cv2.imread(inputFile, -cv2.IMREAD_ANYDEPTH).astype(np.uint8)
+    img = cv2.imread(inputFile, cv2.IMREAD_LOAD_GDAL).astype(np.uint8)
     outputFile = os.path.join(outputDir, currentFilename)
     cv2.imwrite(outputFile, img)