Really helpful to see what the original sample images look like. Personally, I'd like to see more bias toward B/W (IR lit) images, but obviously it needs to have both. Are you aiming for a smaller image set this time around? I presume that would help out those of us whose systems struggle a bit with the current dark.pt when running it on a CPU only Deepstack setup.
I started training the new dark model. I tried to train it local with my RTX 3060 Ti but I could not get it to work. I think I broken it when I was make DeepStackALPR. So i figured I will give Google Colab a try and train the model on their servers, sofar looks to be going well. I should have the model finished sometime late tomorrow or early the next day.
Training still going, it should be done in about 16 hours. The training crashed when I was getting ready to go to sleep using the free version of Colab and would not let me restart the training so I bought a one month subscription to Colab Pro, the cost was only $9.99.
With the new model the accuracy should be better than the original dark models and as fast as my other models
The training looks to still be working (results.txt is being updated), it looks like Colab can read and write to my Google Drive, just I can not upload or download any files
Mike, sorry if this is a stupid or already answered question, but are you always going to store your latest versions of your models on the very first post of this thread?
-Steve
PS. Great work and I think I can speak for most of us DS fans, THANKS!
I just uploaded the new dark model in the first post. This new dark model is faster and better object detection then the original dark model also uses less memory, enjoy.
Let me know how the model works.
I figured out how to download the images that DeepStack used to train the object model. With these images I can start to make the community custom DeepStack model, below are the steps to create the custom model. The first step would be to create a new DeepStack custom model using the same image...
I just uploaded the new dark model in the first post. This new dark model is faster and better object detection then the original dark model also uses less memory, enjoy.
Let me know how the model works.
I figured out how to download the images that DeepStack used to train the object model. With these images I can start to make the community custom DeepStack model, below are the steps to create the custom model. The first step would be to create a new DeepStack custom model using the same image...
It's looking good so far on the cams I loaded it one even in this short of time. It appears fast and highly accurate. You have once again outdone yourself!
Wow, the new dark model is much smaller than the old one. Loading it now. I run combined and dark after sunset and overnight, combined only during the day.
If you are using combined all you need is (objects:0,combined) for daytime and (objects:0,dark,combined) for nightime all of the new animals are in the combined model. In the next several days I am going to add the dark images to the new models so the new models will not need dark any more.
Thank you for this change. That is the one thing that drove me crazy when using the original dark model. I like the "person" or "Person" to differentiate the two running models.
Next I am going to try making a new General Model V2.0. This model I will merge the general images with the dark images and make one model that would only have one label for each item. So no more person & Person. Also I am going to add cat & dog to the model. This model might take days to train because there are 10x more images than when I trained the dark model and this took about 24 hours to train
Labels for General V2.0:
bicycle, bus, car, cat, dog, motorcycle, person, truck