How to train and deploy YOLO Segmentation with cityscapes?

[NJUSTghw]

demo_cityscapes.sh: 1: demo_cityscapes.sh: ./examples/yolo/yolo_detect: not found

[eric612]

1. Make lmdb
2. Use these models to train

However , it is not complete tutorial , you may meet some problems , it is welcome to ask question here.

[kkjong]

Hi, @eric612 I tried to train and deploy YOLO segmentation with cityscapes. I created lmdb with leftImg8bit images, xml annotations and binary label images. When I run mobilenet_yolov3_train.prototxt with pre-trained model you uploaded, I found mismatch num_output in conv26. In test.prototxt, conv26 has num_out : 19, but, in train.prototxt, conv26 has num_out : 3. So, I chagned train.prototxt num_out : 19 to num_out : 3 in conv26, I succeded to train .

But, demo was something wrong. I upload my result:

I think object detection is working well, But segmentation has some problem.

[eric612]

@kkjong If you want to train models , recommend to use imagenet pre-trained. And the channel numbers was based on your lablemap setting

But here has a problem , the deploy model which I gave was 3 class segmentation , however , the default label map setting and trainning prototxt was 19 class segmentation , you need modify this problems , for example :

• Change label map setting and modify trainning prototxt 'conv26' channel 19 to 3
• Or change deploy prototxt 'conv26' channel to 19