olafenwamoses / deepstack_exdark Goto Github PK

Is there coco or yolo format Exdark-dataset?

Not sure if this is an issue or feature request but noticed that the class labels of this model dont match the default model. Specifically, ExDark uses "person" vs "People" and "motorcycle" vs "Motorbike". There is also a capitalisation difference in the class names. This makes it slightly more complicated to configure client applications (e.g. Blue Iris) to filter in/out classes of objects.

I imagine that "normalising" data could be a challenge as more custom models appear but it could also be a real advantage of deepstack if possible.

Do you have the original annotated images and would you be willing to publish or share them?

The YOLOv5x model is being a bit slow for my use case. I would like to try to optimize this data set for my needs, but would rather not have to re-annotate the original exdark set if the work has already been done.

Thanks

Hello @OlafenwaMoses !

First: Thank you for your work on this!!

Now, I just replaced the standard deepstack model with yours, and the speed at which my machine is processing each frame is about half against standard deepstack model. That is: It takes almost twice the time to inspect a video frame as before.

Is this correct ?

On the other hand: it detects People (which is the only object I am interested in) with about twice the certainity, when compared against vanilla deepstack model. Nice !!

Thx again!