Merge pull request #18 from matlab-deep-learning/bugfix/selectStrongestOrder/17

Fixes issue with ordering of boxes and landmarks

Author: justinpinkney

code/mtcnn/+mtcnn/Detector.m

@@ -126,7 +126,7 @@
             landmarks = cat(3, x, y);
             landmarks(probs(:, 2) < obj.ConfidenceThresholds(3), :, :) = [];
 
-            [scores, bboxes] = obj.processOutputs(probs, correction, bboxes, 3);
+            [scores, bboxes, landmarks] = obj.processOutputs(probs, correction, bboxes, 3, landmarks);
 
             % Gather and cast the outputs
             bboxes= gather(double(bboxes));
@@ -151,17 +151,20 @@
             cropped = mtcnn.util.cropImage(im, bboxes, outputSize);
         end
 
-        function [scores, bboxes] = ...
-                processOutputs(obj, probs, correction, bboxes, netIdx)
+        function [scores, bboxes, landmarks] = ...
+                processOutputs(obj, probs, correction, bboxes, netIdx, landmarks)
             % processOutputs    Post-process the output values.
             faceProbs = probs(:, 2);
             bboxes = mtcnn.util.applyCorrection(bboxes, correction);
             bboxes(faceProbs < obj.ConfidenceThresholds(netIdx), :) = [];
-            scores = faceProbs(faceProbs > obj.ConfidenceThresholds(netIdx));
+            scores = faceProbs(faceProbs >= obj.ConfidenceThresholds(netIdx));
             if ~isempty(scores) 
-                [bboxes, ~] = selectStrongestBbox(gather(bboxes), scores, ...
+                [bboxes, scores, index] = selectStrongestBbox(gather(bboxes), scores, ...
                                 "RatioType", "Min", ...
                                 "OverlapThreshold", obj.NmsThresholds(netIdx));
+                if netIdx == 3
+                    landmarks = landmarks(index, :, :);
+                end
             end
         end
 

test/+tests/DetectorTest.m

@@ -131,4 +131,4 @@ function testGpuDetect(test, imageTypeConversion, useDagNet)
             test.verifyEqual(landmarks, test.Reference.landmarks, "RelTol", 1e-1);
         end
     end
-    end
+end

test/+tests/RegressionTest.m

@@ -0,0 +1,29 @@
+classdef RegressionTest < matlab.unittest.TestCase
+    % Test cases for known bugs that have been fixed
+    
+    % Copyright 2020 The MathWorks, Inc.
+
+    methods (Test)
+        function testSelectStrongestBug(test)
+            % GitHub issue #17
+            im = imread("visionteam1.jpg");
+
+            [bboxes, scores, landmarks] = mtcnn.detectFaces(im, "ConfidenceThresholds", repmat(0.01, [3, 1]));
+            for iBox = 1:size(bboxes, 1)
+                test.assertInBox(landmarks(iBox, :, :), bboxes(iBox, :));
+            end
+        end
+    end
+    
+    methods
+        function assertInBox(test, landmarks, box)
+            % check that all landmarks are within the bounding box
+            tf = all(inpolygon(landmarks(1, :, 1), ...
+                            landmarks(1, :, 2), ...
+                            [box(1), box(1) + box(3), box(1) + box(3), box(1)], ...
+                            [box(2), box(2), box(2) + box(4), box(2) + box(4)]));
+            test.assertTrue(tf, "Landmarks should all be inside bounding box");
+        end
+    end
+
+end