search

eric612 / MobileNet-YOLO

A caffe implementation of MobileNet-YOLO detection network
Other
865 stars 442 forks source link

Hi, have you compared your mobileNet-yolov3 with the original yolov3-tiny in speed and precision? #13

Closed zhangjinsong3 closed 5 years ago

zhangjinsong3 commented 6 years ago

As titled, I want to train my data in yolov3 and i find it a little slow of yolov3. Thus, I trained yolov3-tiny but it works not very well. So I want to know if the version of yolov3 with the mobilenet backend perform better than yolov3-tiny?

eric612 commented 6 years ago

Not yet , but I will add coco training result and compare with original yolov3 (also include tiny version) in the future.

zhangjinsong3 commented 5 years ago

@eric612 Thanks for your reply. And have u planned to train an mobileNet_v2 based yolov3?

zhangjinsong3 commented 5 years ago

@eric612 Another question.In your mobilenet_yolov3_train.prototxt, the Yolov3Loss3 layer has a param of side:26,why not 52 in 3rd yolo layer as the darknet is ?

eric612 commented 5 years ago

@zhangjinsong3

  1. I have been try Mnasnet model (better than mobilenet-v2) , the accuracy was not good . Because I can't find a good pre-train weights on imagenet (caffe) . Currently, I don't have so much time (this is my personal project) to improve performance , contribution is welcome
  2. In my test , concat. the 5th convolution layer(52x52) not much help the accuracy , but cost mass computation.
zhangjinsong3 commented 5 years ago

@eric612 Thanks very much for your awesome work！

eric612 commented 5 years ago

Update mobilenet-yolov3 coco mAP