Update AP calculation (#4260)

* Update AP calculation

* Cleanup

* Remove original

val.py

@@ -50,26 +50,27 @@ def save_one_json(predn, jdict, path, class_map):
                       'score': round(p[4], 5)})
 
 
-def process_batch(predictions, labels, iouv):
+def process_batch(detections, labels, iouv):
-    # Evaluate 1 batch of predictions
+    """
-    correct = torch.zeros(predictions.shape[0], len(iouv), dtype=torch.bool, device=iouv.device)
+    Return correct predictions matrix. Both sets of boxes are in (x1, y1, x2, y2) format.
-    detected = []  # label indices
+    Arguments:
-    tcls, pcls = labels[:, 0], predictions[:, 5]
+        detections (Array[N, 6]), x1, y1, x2, y2, conf, class
-    nl = labels.shape[0]  # number of labels
+        labels (Array[M, 5]), class, x1, y1, x2, y2
-    for cls in torch.unique(tcls):
+    Returns:
-        ti = (cls == tcls).nonzero().view(-1)  # label indices
+        correct (Array[N, 10]), for 10 IoU levels
-        pi = (cls == pcls).nonzero().view(-1)  # prediction indices
+    """
-        if pi.shape[0]:  # find detections
+    correct = torch.zeros(detections.shape[0], iouv.shape[0], dtype=torch.bool, device=iouv.device)
-            ious, i = box_iou(predictions[pi, 0:4], labels[ti, 1:5]).max(1)  # best ious, indices
+    iou = box_iou(labels[:, 1:], detections[:, :4])
-            detected_set = set()
+    x = torch.where((iou >= iouv[0]) & (labels[:, 0:1] == detections[:, 5]))  # IoU above threshold and classes match
-            for j in (ious > iouv[0]).nonzero():
+    if x[0].shape[0]:
-                d = ti[i[j]]  # detected label
+        matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detection, iou]
-                if d.item() not in detected_set:
+        if x[0].shape[0] > 1:
-                    detected_set.add(d.item())
+            matches = matches[matches[:, 2].argsort()[::-1]]
-                    detected.append(d)  # append detections
+            matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
-                    correct[pi[j]] = ious[j] > iouv  # iou_thres is 1xn
+            # matches = matches[matches[:, 2].argsort()[::-1]]
-                    if len(detected) == nl:  # all labels already located in image
+            matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
-                        break
+        matches = torch.Tensor(matches).to(iouv.device)
+        correct[matches[:, 1].long()] = matches[:, 2:3] >= iouv
     return correct