Primeira implementação de métricas de detecção

examples.py

@@ -3,20 +3,18 @@
 
 Requires:
     opencv-python, numpy
-
-TODO:
-    Detection example
 """
 import cv2
 import numpy as np
 
-from lapixdl.evaluation.evaluate import evaluate_segmentation, evaluate_classification
+from lapixdl.evaluation.evaluate import evaluate_segmentation, evaluate_classification, evaluate_detection
 from lapixdl.evaluation.model import BBox, Classification
 
 
 def main():
     # Model evaluation examples
     evaluate_segmentation_example()
+    evaluate_detection_example()
     evaluate_classification_example()
 
 
@@ -73,6 +71,34 @@ def evaluate_segmentation_example():
     # Shows confusion matrix for class `a`
     metrics.by_class[0].show_confusion_matrix()
 
+def evaluate_detection_example():
+    # Class names
+    classes = ['kite', 'person']
+
+    # Image shape
+    mask_shape = (480, 640)
+
+    # Creating fake data    
+    gt_bbox_1 = BBox(10, 10, 10, 10, 0)
+    pred_bbox_1 = BBox(10, 10, 10, 10, 0)
+
+    gt_bbox_2 = BBox(110, 110, 320, 280, 1)
+    pred_bbox_2 = BBox(70, 50, 240, 220, 1)
+
+    # Creating data suplier iterator
+    # It is not necessary here, but it's useful if you want to yield data
+    # from the disk i.e. from a Pytorch DataLoader
+    it_gt_masks = identity_iterator([gt_bbox_1, gt_bbox_2])
+    it_pred_masks = identity_iterator([pred_bbox_1, pred_bbox_2])
+
+    # Calculates and shows metrics
+    metrics = evaluate_detection(it_gt_masks, it_pred_masks, classes)
+
+    # Shows confusion matrix and returns its Figure and Axes
+    fig, axes = metrics.show_confusion_matrix()
+
+    # Shows confusion matrix for class `a`
+    metrics.by_class[0].show_confusion_matrix()
 
 def evaluate_classification_example():
     # Class names

lapixdl/evaluation/evaluate.py

@@ -90,7 +90,8 @@ def __merge_detection_metrics(acc: DetectionMetrics,
 def evaluate_detection(gt_bboxes: Iterable[List[BBox]],
                        pred_bboxes: Iterable[List[BBox]],
                        classes: List[str],
-                       iou_threshold: float = .5) -> DetectionMetrics:
+                       iou_threshold: float = .5,
+                       undetected_cls_name: str = "_undetected_") -> DetectionMetrics:
     """Evaluates detection predictions
 
     The iterables should return the bboxes of one image per iteration 
@@ -102,19 +103,37 @@ def evaluate_detection(gt_bboxes: Iterable[List[BBox]],
         pred_bboxes (Iterable[List[BBox]]): Predicted classifications.
         classes (List[str]): Class names.
         iou_threshold (float): Minimum IoU threshold to consider a detection as a True Positive.
+        undetected_cls_name (str): Name to be used for undetected instances class. Defaults to "_undetected_".
 
     Returns:
         DetectionMetrics: Detection metrics.
     """
 
-    confusion_matrix = np.zeros((len(classes), len(classes)), np.int)
+    confusion_matrix = np.zeros((len(classes) + 1, len(classes) + 1), np.int)
+    undetected_idx = len(classes)
 
     for (curr_gt_bboxes, curr_pred_bboxes) in tqdm(zip(gt_bboxes, pred_bboxes), unit=' samples'):
         pairwise_bbox_ious = calculate_pairwise_bbox_ious(
             curr_gt_bboxes, curr_pred_bboxes)
-        #TODO: Evaluate IoU results
 
-    metrics = DetectionMetrics(classes, confusion_matrix)
+        no_hit_idxs = set(range(0, len(curr_pred_bboxes)))
+
+        for i, gt_ious in enumerate(pairwise_bbox_ious):
+            gt_cls_idx = curr_gt_bboxes[i].cls
+
+            hits_idxs = [i for i, iou in enumerate(gt_ious) if iou >= iou_threshold]
+            no_hit_idxs.difference_update(hits_idxs)
+
+            if len(hits_idxs) == 0:
+                confusion_matrix[undetected_idx, gt_cls_idx] += 1
+            else: 
+                for hit_idx in hits_idxs:
+                    confusion_matrix[curr_pred_bboxes[hit_idx].cls, gt_cls_idx] += 1
+
+        for no_hit_idx in no_hit_idxs:
+            confusion_matrix[curr_pred_bboxes[no_hit_idx].cls, undetected_idx] += 1
+
+    metrics = DetectionMetrics(classes + [undetected_cls_name], confusion_matrix)
 
     print(metrics)
 
@@ -152,7 +171,7 @@ def calculate_bbox_iou(bbox_a: BBox, bbox_b: BBox) -> float:
     Returns:
         float: IoU between the two bboxes.
     """
-    
+
     # Gets each box upper left and bottom right coordinates
     (upr_lft_x_a, upr_lft_y_a) = bbox_a.upper_left_point
     (btm_rgt_x_a, btm_rgt_y_a) = bbox_a.bottom_right_point

lapixdl/evaluation/model.py

@@ -391,16 +391,17 @@ def avg_precision(self):
 
 
 class DetectionMetrics(ClassificationMetrics):
-    """Multiclass instance-based classification metrics
+    """Multiclass detection metrics
 
     Attributes:
         classes (List[str]): Class names.
-        confusion_matrix (List[List[int]]): Confusion matrix of all the classes.
+        confusion_matrix (List[List[int]]): Confusion matrix of all the classes. 
+        The last colunm and line must correspond to the "undetected" class.
     """
 
     def __init__(self, classes: List[str], confusion_matrix: List[List[int]] = []):
         super().__init__(classes, confusion_matrix)
-        self._by_class = [BinarySegmentationMetrics(x) for x in self.by_class]
+        self._by_class = [BinaryDetectionMetrics(x) for x in self.by_class]
 
     @property
     def avg_iou(self):

tests/evaluation/evaluate/test_evaluate_detection.py

@@ -0,0 +1,11 @@
+import pytest
+import numpy as np
+
+from lapixdl.evaluation.evaluate import evaluate_segmentation
+from lapixdl.evaluation.model import BBox
+from ... import utils
+
+
+def test_evaluation_detection_metrics():
+    #TODO: Implement detection metrics tests
+    pass