DeepStack ExDark - Detect objects in dark/night images and videos

As part of the effort to provide a collection of sample custom DeepStack models trained on popular and custom datasets, we are happy to announce the release of the DeepStack_ExDark custom model for detecting people and 11 other common objects in dark/night images and videos. It is a YOLOv5x custom model trained on the Exdark dataset. This model address issue of missed detections, poor accuracy and mis-classification of objects in dark/night images.

You can put DeepStack_ExDark custom model to test by visiting the linked below.

dark5_detected

Also, you can find more custom models by our team and by others in the community on our documentation linked below.

https://docs.deepstack.cc/custom-models-samples/

4 Likes

@OlafenwaMoses this is excellent, thank you for your continued work.

In the future, do you plan to include this as part of the standard DS release?

1 Like

@aesterling thanks for the kind words. :slightly_smiling_face:

We don’t plan to ship DeepStack_ExDark custom model as part of the standard APIs, reason being that we will be publishing lots of custom models as time goes by, and can’t always add a lot of these model to the Core API to avoid shipping unnecessarily large packages for DeepStack.

You can always use the model via the Custom Model API across all the versions of DeepStack.

2 Likes

Thanks for your hard work. Would it be possible to provide a version with all the models for us folks who do not mind the large size? A separate link than the default. I dont know enough to tinker with it myself.
Best

Love that the DeepStack folks are planning on creating / releasing custom model packs. One model pack that I think would be welcomed is a small animal pack identifying critters like raccoon, fox, weasel, owl, hawk - predatory type animals that typically one does not want in their back yard :slight_smile:

I’m excited to see what model packs are made available… will these be plug-n-play? As in drop the model in the custom model directory and you’re good to go? :+1:t2::+1:t2:

1 Like

@linux It’s not feasible for us to ship all the models in a separate version. Our goal is to have 100s of custom models available to use freely in the DeepStack community in the next 1-2 years and thousands later on.

Are you having issues deploying and testing the custom model? What are the issues?

@dirk6665 while we keep putting effort to release models with more objects covered, there isn’t any guarantee these specific objects will be covered.

My advice is to take the effort to

It might take some time and effort, but the result is always satisfactory once you have your working model.

I have had a look at the “dark” model and the docs suggest this has a different endpoint to the standard “day” model. I think this means that I cant “just drop it in” to Blue Iris implementation and have both night and day detection. Is that correct ? Or is there a way to combine the two models to get best of both worlds ?

Really appreciate the efforts on this as it has made a step-change in motion detection in security camera domain.

1 Like

Moses, thanks for the reply. I have not tried deploying a custom model but I would like to, for example combine two models at once. The standard model and the Exdark. I dont quite understand how to do that. Do i need to run two instances at once? I am using DS with blue iris. Blue iris allow me to specify a folder that contain my models but I dont know if I can simply dump a bunch of random models in that folder.
Best regards,

1 Like

@linux You can deploy any number of custom models and the standard APIs in DeepStack in a single instance.

You can check out a sample of running multiple APIs here.

https://docs.deepstack.cc/using-deepstack-with-nvidia-gpus/index.html#run-with-all-apis

1 Like

Thanks for the response. After some investigation with BI support, I have both the default and ExDark models working together. Many thanks.

1 Like

Moses, do you make the yolo annotation files available for the deepstack models ? I dont see them on the repo listed above.

No, the annotated dataset is not included in the repository.

Ahh, thought the models were open-source as well. Sorry.

The trained YOLOv5 model is open source. It is the annotated dataset that it is not open source.

1 Like

Hi @PeteB, how were you able to have the custom models and the standard APIs working in BI? Thank You!!!

@keshav, you create a custom model directory somewhere on your BI server and copy in the .pt file from the repo link above. You need to tell BI about the custom directory in the BI Global Setting’s AI tab. This will enable both the default model as well as the ExDark one. Make sure you are on a recent BI version.

If you want to confirm it is working then you can start a DeepStack instance from the command prompt before starting the BI service, e.g.something like deepstack --VISION-DETECTION True --MODELSTORE-DETECTION "C:\path-to-custom-model" --PORT 82. You should then see log messages that show both models being used as below.


Note that the ExDark uses slightly different object labels than the default model, e.g. “person” vs “People” so you will need to modify the camera AI settings to include/exclude the required labels. Note also that this significantly increases the AI processing time. I suspect because the dark model is yolov5x whereas the default model is yolov5m.

2 Likes