I am using DeepStackAi with Blue Iris (a video security software).
I have configured it with a test camera.
The camera I use is one that has quiet good night vision (IPC-HDW5231R-Z).
I am only identifying persons in the generated pictures (-e VISION-DETECTION=True in the docker command). I am using the CPU version.
Day time detection seems to work as it should. I did not get any false alerts during my testing.
Nighttime is very different - it doesn’t detect any persons at all.
I am pretty sure all is configured OK since day time works without any issues but not sure why it doesn’t detect at night.
I can try and tweak the night time picture - more contrast and such.
But I am wondering if there is a better way? Like to “train” the model by feeding it some of the nighttime pictured generated by Blue Iris?
Hello @MnM, thanks for bringing this up.
You can run DeepStack in high accuracy model by adding
-e MODE=HIgh . This would run slower but would detect better.
Also, can you share sample images from the test camera, so we can investigate this better?
@john are there object detection models trained on images from night vision (IR cameras like the IPC-HDW5231R-Z)? I am also interested in true thermal imaging (FLIR) cameras
@robmarkcole, deepstack supports detection at night. Below are two sample night images i just tried with DeepStack.
On thermal images, datasets are highly limited in that domain, however, it is something we shall still work on.
I expect the pictures that @MnM are not of such high resolution and contrast as these example images, but it is very interesting that a silhouette works so well! Suggests that some pre-processing of the IR images might improve accuracy, perhaps by applying thresholding?
My issue turned out to be Blue Iris related in the end.
Nighttime persons are detected fine by Deepstack.
How did you manage to tune Deepstack in order to have good nightime image detection? I have False Negatives always with IR b/w images and I have a good IR camera.
During daytime works really great.