Adding support for evaluating keypoints

fiftyone/core/patches.py

@@ -23,7 +23,7 @@
 import fiftyone.core.view as fov
 
 
-_PATCHES_TYPES = (fol.Detections, fol.Polylines)
+_PATCHES_TYPES = (fol.Detections, fol.Polylines, fol.Keypoints)
 _NO_MATCH_ID = ""
 
 

fiftyone/utils/eval/detection.py

@@ -130,7 +130,7 @@ def evaluate_detections(
     fov.validate_collection_label_fields(
         samples,
         (pred_field, gt_field),
-        (fol.Detections, fol.Polylines, fol.TemporalDetections),
+        (fol.Detections, fol.Polylines, fol.TemporalDetections, fol.Keypoints),
         same_type=True,
     )
 

fiftyone/utils/iou.py

@@ -17,6 +17,8 @@
 import fiftyone.core.utils as fou
 import fiftyone.core.validation as fov
 
+from scipy.spatial import distance
+
 from .utils3d import compute_cuboid_iou as _compute_cuboid_iou
 
 sg = fou.lazy_import("shapely.geometry")
@@ -89,6 +91,12 @@ def compute_ious(
             preds, gts, error_level, iscrowd=iscrowd, classwise=classwise
         )
 
+
+    if isinstance(preds[0], fol.Keypoints):
+            return _compute_keypoint_similarity(
+                preds, gts, iscrowd=iscrowd, classwise=classwise
+            )
+
     if use_masks:
         # @todo when tolerance is None, consider using dense masks rather than
         # polygonal approximations?
@@ -556,6 +564,28 @@ def _compute_polyline_ious(
         return ious
 
 
+def _compute_keypoint_similarity(preds, gts, error_level, iscrowd=None, classwise=False, gt_crowds=None):
+    #calculate euclidean distance of each keypoint
+    sim_score = np.zeros((len(preds), len(gts)))
+
+    for j, gt in enumerate(gts):
+        for i, pt in enumerate(preds):
+            gt_points=[]
+            pt_points=[]
+            for x in gt['points']:
+                gt_points.append(x)
+            for y in pt['points']:
+                pt_points.append(y)
+
+            gt_points=sum(gt_points, [])
+
+            pt_points=sum(pt_points, [])
+
+            sim_score[i, j] = 1/(1 + distance.euclidean(tuple(gt_points),tuple(pt_points)))
+
+    return sim_score
+
+
 def _compute_mask_ious(
     preds, gts, tolerance, error_level, iscrowd=None, classwise=False
 ):