Jump to content
Sign in to follow this  
marcbot123

Using mybot to generate base dataset for deeplearning ?

Recommended Posts

[I probably didn’t post in the right section of the forum, please move this in the right section]

Hello,

 

How hard would it be to use/modify the code to make it ''annotate'' bases ?

 

The goal would be to add bounding boxes arround buildings and creating xml/json/txt files of all the bounding boxes assosiated with a building. I have been doing it for 200 images and it takes alot of time and my dataset is only th11-th12. I have been training an object detection deeplearning algorithm on this dataset and the results are really good, so if we could use mybot to ''annotates'' alot more bases, the model would be insane at detection (can do up to 150 images per second on a gtx1080, and close to 15 images per second on cpu).

 

Is there someone who would be willing to help on this ?

 

Here is what I am talking about the bounding boxes (doing it by hand) :

 

f52IadF.jpg

 

 

Edited by marcbot123

Share this post


Link to post
Share on other sites

Right now, the deeplearning part is not my problem, I need to get the dataset and would need someone who knows how to tweek the bot engine to make it label things with the already built in image detection to speed-up the labeling a little and add lower th bases to my dataset. 

 

Keras is great, but I’ll be using Darkflow which is a Tensorflow version of Darknet (Yolo) 

 

thanks

 

Edited by marcbot123

Share this post


Link to post
Share on other sites

 

 

 

 

Trained this object detection model to help create metadata about base buildings. This requires no mod since it only takes the game image as input and will output the coordinates of the buildings it detects (at 27 fps on a 1080ti).

 

This first version is trained on 300 bases dataset I manually annotated (took over 2 month to get this done on my free time). Next step is to use this model to auto-label some more bases and augment the dataset to get a better model. This currently achives 80% mAP at 0.5 and uses yolov2 (Darkflow) for those who know a bit about deeplearning.

 

The plan is to create a SaaS where people will be able to upload a video of how they attacked a certain base. The video will be stocked and a metadata file of the buildings will be created. The person would then be able to upload the image of a base he wants to attack and search for "most similar" bases in his own already uploaded videos. This will be really useful for CWL since you only get 1 swing per base. This service could be shared inside a clan for example.

 

I am posting here to see if people have an interest in it. I am a data scientist, I don't know much about frontend and SaaS stuff, so if someone is interested to work with me for this part, I am sure we could work something out.

 

Thats pretty much it. Forgive my English, its not my first language.

Thanks

Edited by marcbot123

Share this post


Link to post
Share on other sites

Do you think it possible to similarly train on bases(building positions) and army compositions used for the 3 star(goho, lalo)? So give an image of the base in the same fashion you are doing and have the model just predict the best general strategy if there is not a video on it?

 

From there the user could take the suggestion and work around that for their attack. There will not be a definite right or wrong, but I wonder if it could help on very tricky unseen bases to paint a general guide line.

Share this post


Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Sign in to follow this  

×
×
  • Create New...