Improving detections by object identity

[illigtr]

If I am not mistaken, open AI vision is being used to detect lines in captured images. Can detection improve by using a better or trained model with open AI vision? Clearly most, if not all, aircraft trails have flashing navigation lights. Satellites pose a problem, as they do not have nav lights. I'd be happy just removing 95% of the aircraft, leaving only satellites and meteoroids to analyze.

My typical daily routine is to review the nightly images with detection mode on. Then using only the left arrow button I can review all detections historically. This works well from my desktop, but is a pain on a tablet or smartphone, as no arrow keys exist or if they do, it force the soft keyboard to overlay the already busy screen. Could a next/previous button be placed on the bottom on the image, or does some equivalent functionality exist when reviewing detection images from a tablet/smartphone?

[aaronwmorris]

A model is not being used for meteor detection. It is far more simple... it is just line detection using OpenCV.

The Gallery in indi-allsky should have relatively decent mobile integration. It supports swiping to cycle through images. Pinch to zoom works as well. I probably do need to add mobile event integration for the Image Viewer page.

[illigtr]

Yes, I was afraid of that. I started looking at the code for detection after writing this request. I will have to try again the swiping method while the detection filter is on. Zoom does work well. Perhaps due to my use of Firefox on the portable devices? Will retest.

[illigtr]

I just tried from my phone, using Edge and Firefox, and swiping has no effect.Perhaps I am not swiping from the correct screen?I select Media, Images. I then turn on the detections switch. Swiping doesn't seem to work. If I click the image I go full screen, but swipe does not appear to work there as well.Stumped.On Dec 27, 2025 19:00, Aaron W Morris ***@***.***> wrote: A model is not being used for meteor detection. It is far more simple... it is just line detection using OpenCV. The Gallery in indi-allsky should have relatively decent mobile integration. It supports swiping to cycle through images. Pinch to zoom works as well. I probably do need to add mobile event integration for the Image Viewer page. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>

[aaronwmorris]

I am adding the enhancement tag for gesture support for the Image viewer. There appear to be some javascript libraries for this.

[illigtr]

Oh...wait...the gallery...missed that...On Dec 27, 2025 19:42, Tony ILLIG ***@***.***> wrote:I just tried from my phone, using Edge and Firefox, and swiping has no effect.Perhaps I am not swiping from the correct screen?I select Media, Images. I then turn on the detections switch. Swiping doesn't seem to work. If I click the image I go full screen, but swipe does not appear to work there as well.Stumped.On Dec 27, 2025 19:00, Aaron W Morris ***@***.***> wrote: A model is not being used for meteor detection. It is far more simple... it is just line detection using OpenCV. The Gallery in indi-allsky should have relatively decent mobile integration. It supports swiping to cycle through images. Pinch to zoom works as well. I probably do need to add mobile event integration for the Image Viewer page. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>

[illigtr]

I am considering training a model with YOLOv8 for Python to distinguish meteoroids from aircraft. I probably have enough images for both to get a somewhat decent model. I presume I could include the code block somewhere within and/or replace the detectLines.py for testing. Have you heard of anyone else attempting something similar?

[aaronwmorris]

I have not heard of anyone trying to generate a model to differentiate between meteors and aircraft. Several people have used models to determine cloudiness/weather. The models are generally trained on the whole image where the image is scaled down significantly.

How would you train a model? Would you use the whole image or crop just the meteors/aircraft for training?

[illigtr]

The whole images have to be broken up with bounding boxes and labels for training. However, once the model is trained, it can process the entire image at once.