Variable Image size as Input to train own dataset.

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).

MIT License

1.73k stars 784 forks source link

Variable Image size as Input to train own dataset. #251

Closed R1234A closed 6 years ago

R1234A commented 6 years ago

I am using "keras-yolo2" for my dataset in which the image size is variable means 1280*720 px. I want to ask, in the config.json file the input_size have only one dimension. So how can I give two dimensions there??

Can I give None in input_size?? So, that it can automatically takes my input image size??

Please reply asap.

experiencor commented 6 years ago

Nope. Variable input size is not currently implemented in this repo of mine. Try keras-yolo3 instead.

ShilpaSangappa commented 6 years ago

Does that mean pascal VOC dataset cannot be used for training, as the images in it are of different sizes?

jmaity commented 6 years ago

@experiencor Do I need to resize every image before annotation? I'm bit confuse in this regard as i trained the model using my own data set and no box was detected. Please help me with the 'Input Size' variable. Thanks in advance.