This is the result of OpenCV YOLOv2
While this is the result of using darknet YOLOv2
May I know why opencv YOLOv2 is different from darknet's? Should both of the results are different? If I'm wrong in any way please do correct me.
cfg = "yolo.cfg" weights = "yolo.weights" names = "coco.names" threshold = 0.24
Please add more details when you refer a public models. I've found at least 3 YOLOv2 at https://pjreddie.com/darknet/yolo/ and they aren't similar. I hope YOLOv2 trained on COCO dataset with input size 608x608 is used. Am I right? OpenCV's sample works with 416x416 images for now so you need to modify a source code a bit:
from
Mat inputBlob = blobFromImage(frame, 1 / 255.F, Size(416, 416), Scalar(), true, false);
to
Mat inputBlob = blobFromImage(frame, 1 / 255.F, Size(608, 608), Scalar(), true, false);
One more thing that corresponding .cfg file has a confidence threshold inside:
thresh = .6
I don't know why Darknet ignore it but OpenCV consider this value to threshold detections with a low confidence. So you have to modify it too. In example,
thresh = .1
The result:
Hidden objects has confidences about 0.4, 0.29, 0.37.
Asked: 2018-01-15 12:59:10 -0600
Seen: 2,943 times
Last updated: Jan 16 '18
Afaik OpenCV has YOLOv1 because it does not have these return layers that YOLOv2 needs and hence the lower accuracy?
https://github.com/pjreddie/darknet/i...
as mentioned here, AlexeyAB states that: Now Darknet Yolo v2 is added to the OpenCV: opencv/opencv#9705 You can use it from master-branch or since OpenCV 3.4.0 will be released.
so which one is right? is OpenCV's v1 or v2?