Is it possible to run deepstack detection at image-size 1280?

So I’m training my custom model to identify my cats on my outdoor CCTV cameras. After reading the yolov5 training tips, I believe running img size 1280 would be beneficial for my use case. The cameras are 4k and mounted quite high to cover more area, this results in the cats that i want to detect are a very small part of the total picture.

As the FAQ mentions, it’s recommended to run the detection at the same image size as the training, my question is therefore if there’s any way to run deepstack detection with the same higher image size? (currently running windows 10 native GPU).

Feeding Deepstack larger images increases processing time, is this because the actual size is processed fully (in that case i can adjust the input size), or are all pictures resized to 640px anyway when processed, and the processing time increase is basically just an increase in time the resizing operation to process the larger input image?

I’m sorry if the question is a bit all over the place.

TLDR:

  1. Will deepstack process the input image at its sent resolution or will it resize it internally before running the detection model?
  2. If it is resized internally, is there a way to set the parameter -img 1280 in the yolov5 detection model?
2 Likes

@MissMusic

  1. DeepStack resizes the input image based on the version and configuration used when running the server. Find the various resolutions used here. These values are defined for optimal results with accuracy and performance/compute usage.

  2. There isn’t an option to specify the image resolution to be used by DeepStack yet.

Thank you!

Now i know what settings to use when training my custom model, from what I’ve read to get the best accuracy you should always train and detect at the same settings.

I’m running on GPU with high, so yolov5m with its default image-size (640) should give the best results.

1 Like

In our FAQ, we never mentioned you using a specific size for images/frames you send to DeepStack. The resolution of your images is independent of the type of model you train. You can send any image dimension to any model type. DeepStack will automatically handle the resize for you.

Yea i was not referring to the image size sent to deepstack, but what image-size (like 640 or 1280) were used in the detection so i could match my training. I did get the answer though in the link, as the low, medium and high settings seem to be setting this parameter.