About dilation convs

[ChenCong7375]

Dear Alexey, I noticed that you added dilation in conv/lstm,Could you please tell me how could I use it? thank you very much.

[ChenCong7375]

Looking forward to know how to use it in a .cfg file.

[ChenCong7375]

@AlexeyAB How to set dilation parameter in [convolutional] layers? should I make .cfg such as : [convolutional] batch_normalize=1 dilation=1 filters=16 size=3 stride=1 pad=1 activation=leaky ?

[AlexeyAB]

@ChenCong7375 Hi,

• How to use Dilation, look at TridentNet https://github.com/AlexeyAB/darknet/blob/master/cfg/resnet152_trident.cfg More: https://github.com/AlexeyAB/darknet/issues/3363

• How to use LSTM-conv networks, look at Yolov3-tiny-lstm (must be trained on sequential images/frames from video - so in train.txt images should go consistently rather than randomly - one or several sequences): https://github.com/AlexeyAB/darknet/files/3199631/yolo_v3_tiny_pan_lstm.cfg.txt More: https://github.com/AlexeyAB/darknet/issues/3114

[ChenCong7375]

@AlexeyAB Hi, and could I visualize the feature map of any convolutional layer in your project?

[NickiBD]

Hi @AlexeyAB I want to use [dilated_convolutional] header in my .cfgs to help me further for very small object detection. I found an implementation in zzzDavid/Dilated-Convolutional-Layer-Implementation repo (An Implementation of Dilated Convolutional Layer Based on Darknet Architecture) .I tried to integrate it with your repo .However, it doesnot recognize the header yet .I am not sure ,whether it is doable ,how to integrate it and what I did was right .I would be really grateful if you could help me with this .

Many thanks again for all the help .

[AlexeyAB]

@NickiBD Hi,

Are you about this repo? https://github.com/zzzDavid/Dilated-Convolutional-Layer-Implementation

What model do you want to use and how dilated conv layers can help to detect small objects, are there any papers or info?

[NickiBD]

Hi @AlexeyAB Thank you so much for the reply. yes .This is the repo .I want to use it with tiny yolo v3 models and dilation-rate=1 .I tried to use dilation rates( defined in Trident network ) with tiny yolo v3 3l ( I need a light model) but the mAP was zero . I found this paper https://arxiv.org/pdf/1709.00179 so I was thinking I may be able to use this to increase the receptive field to have better accuracy for the small objects. I would be grateful if you could help me .

Thanks again.

[AlexeyAB]

@NickiBD Dilation convolution is already implemented in this Darknet repository https://github.com/AlexeyAB/darknet/issues/3363#issuecomment-502313321

and tested well on TridentNet, comparison: https://github.com/AlexeyAB/darknet/issues/3114#issuecomment-494148968

just look how it is done: https://github.com/AlexeyAB/darknet/blob/master/cfg/resnet152_trident.cfg

[NickiBD]

@AlexeyAB

Thanks for the information and the guidance .I will do so.