Loss metric while Inference .pb file

[kmonachopoulos]

Hello,

Thank you very much for the job that you have done with yolo. I have managed to convert .cfg and .weight to a .pb file (yolo-voc.pb and yolo-voc.meta). For inference I am using :

./flow --pbLoad built_graph/yolo.pb --metaLoad built_graph/yolo.meta --imgdir sample_img/

and the output is :

Forwarding 8 inputs ...
Total time = 1.30014300346s / 8 inps = 6.15316928883 ips
Post processing 8 inputs ...
Total time = 0.138490915298s / 8 inps = 57.7655218955 ips

However, the images that are stored in the out folder that is created are not containing any bounding boxes ...

My final goal is to inference VOC2011 database and extract the total error based on the images in :

VOCdevkit/VOC2011/JPEGImages

Now, in the original paper they are using mAP as an error metric. What exactly is the prediction error above? Is there another way to extract the final error using the flow script ? How can I extract .jpg outputs with the bounding boxes on the objects?

PS. In my case return_predict() returns an empty matrix [] when I am using this .pb file through pbLoad and metaLoad following your example :

from darkflow.net.build import TFNet
import cv2

options = {"pbLoad": "built_graph/yolo-voc.pb", "metaLoad": "built_graph/yolo-voc.meta", "threshold": 0.1}

tfnet = TFNet(options)

imgcv = cv2.imread("./sample_img/sample_dog.jpg")
result = tfnet.return_predict(imgcv)
print(result)

Is there anything wrong in the way I am calling it?

Thanks again.

[macro-dadt]

did you try with all images you trained?

[kmonachopoulos]

It is not about the images, it is about the error metric. I suppose I have to make a script to inference through the VOC.

[Yeaminul]

Hi @kmonachopoulos , Had you been able to make it work? Facing same problem. Any suggestion?