search

xslittlegrass / CarND-Vehicle-Detection

Vehicle detection using YOLO in Keras runs at 21FPS
389 stars 175 forks source link

Different weights file #2

Closed subodh-malgonde closed 7 years ago

subodh-malgonde commented 7 years ago

Hi,

I was having a look at the YOLO V1 website https://pjreddie.com/darknet/yolov1/. It says the tiny-yolo.weights file should be around 103 MB. The link provided by you the readme is about 180 MB. Why are these weight files different?

Also you have provided a reference to https://github.com/sunshineatnoon/Darknet.keras in your utils file. The model architecture in that repository seems to be different than yours, both are supposed to be for YOLO v1. The Darknet.keras repo doesn't have leaky relu after every convolutional layer, but yours does. Why is that so?

Thanks, Subodh

xslittlegrass commented 7 years ago

Hi Subodh,

The 103MB file is for the version 1.1, which has a different architecture. If you see the RunTinyYOLO.py in Darkent.keras, you will see that it does include leaky relus.

subodh-malgonde commented 7 years ago

Thanks!

subodh-malgonde commented 7 years ago

So which is your architecture, 1.0 or something else? Can you give me a link to the research paper for your YOLO architecture?

Figured it out.