Accurary implementation? about yolov3-tf2 HOT 3 CLOSED

I've done so on a custom detection script that i use to choose the checkpoint with better results on my validation dataset. I won't share the code since is very specyfic for my dataset, but the pseudo code would be:

for checkpoint in checkpoints:
-model.load_weights(checkpoint)
-for img, labels in val_dataset:
--boxes, scores, classes, nums = model.predict(img)
--labels2 = labels.copy()
--acc = 0
--for i in range(nums[0]):
---box = boxes[0][i]
---c = classes[0][i]
---for l in labels2:
----gt_box = getBox(l)
----if boxesIntersect(box, gtbox) and sameClass(c, l):
-----acc += 1
-----labels2.remove(l)
-----break
--accuracy = acc / len(labels)

from yolov3-tf2.

merging with #125

from yolov3-tf2.

@DeltaTesting it seems like in your code, one detection (predicted box) would just match the first ground truth box that satisfy the iou and class condition. so if that particular ground truth box is better to be matched with another detection, there would be a miss here. Does your dataset mostly just contains one bounding box, or are those ground truth boxes usually not close with each other?

from yolov3-tf2.