search

mystic123 / tensorflow-yolo-v3

Implementation of YOLO v3 object detector in Tensorflow (TF-Slim)
https://medium.com/@pawekapica_31302/implementing-yolo-v3-in-tensorflow-tf-slim-c3c55ff59dbe
Apache License 2.0
891 stars 353 forks source link

yolo-v3-tiny does not give any prediction #12

Open rohitkumara opened 6 years ago

rohitkumara commented 6 years ago

I have used your demo.py code along with yolo-v3 and yolov3.weights and it predicts the objects in an image very well, but when I run the demo.py code with yolo-v3-tiny and yolo-tiny.weights, the predicted output numpy array is always populated with NaN. I am unable to find a reason why yolo-v3-tiny is giving this output. Could you give a solution or reason as to why this might be happening? I have attached an image of the predicted output being printed to the console. tiny_error

rohitkumara commented 6 years ago

Correction: I used the wrong weights. On using the correct weights ('yolov3-tiny.weights'), I am not getting as much confidence using yolo-v3-tiny as I am getting using yolo-v3. I'm getting 0.99 confidence for a person when using yolo-v3, but i'm only getting 0.37 confidence when I use yolo-v3-tiny network. Could you provide the solution or any reason why? Thanks in advance.

mystic123 commented 6 years ago

Do you get different results using official code from pjreddie's repo (https://github.com/pjreddie/darknet)?

rohitkumara commented 6 years ago

I haven't tried pjreddie's actual repo since my whole project is in Tensorflow and his version (coded using darknet) won't be compatible. I referred to his paper for the theory and the network architecture though. So that's why is used your tutorial and repo. I will try it and let you know the results.

mystic123 commented 6 years ago

Sure, I understand that, but I want to check if the problem is with my code or something else. I am waiting for the results :)

rohitkumara commented 6 years ago

So I tried using pjreddie's code and these are the results I got. For yolo-v3: download.jpeg: Predicted in 27.747483 seconds. traffic light: 70% car: 100% bicycle: 99% person: 100% person: 93% person: 80% person: 78% person: 72% And for yolov3-tiny: download.jpeg: Predicted in 2.408129 seconds. traffic light: 66% car: 92% bicycle: 88% So people are not being detected using yolov3-tiny, by the car and bicycle are being detected with decent accuracy. I also wanted to inquire if you could advise me on getting started with training yolov3-tiny on my own dataset?

mystic123 commented 6 years ago

@surgeonofdeath I updated the code by changing upsampling method to nearest neighbor. Could you please check now if it's working better?

sushant1727 commented 6 years ago

How can I check this code on webcam ?