Considering running something like YOLOv(Something) over all of my photos and I am now just learning that all of the easily downloadable models detect kinda a weird set of things, like look at this https://github.com/ultralytics/ultralytics/blob/d79a7332db43d341c9fd1b1cec0de4596365ea07/ultralytics/cfg/datasets/coco.yaml#L19 Like, yes cool, I want Person, bike, cat, dog. But I don't really need to detect microwave, broccoli, toaster, etc Seems like there are no other good models for "normal things" ?
benjojo
replied 26 Jun 2025 16:08 +0000
in reply to: https://benjojo.co.uk/u/benjojo/h/1gFkzQm2httXYv2G37
hmmm, close, but not quite right Mr YOLOv12
alex@feed.yopp.me
replied 26 Jun 2025 16:23 +0000
in reply to: https://benjojo.co.uk/u/benjojo/h/Hmj1B4h2br84YDrZ28
benjojo
replied 26 Jun 2025 16:35 +0000
in reply to: https://feed.yopp.me/users/alex/statuses/114750589231217485
@alex Seems to be the same stuff dataset wise for their pre trained models, I guess I now understand the moat that the other companies have on this sector
loop0@freeradical.zo..
replied 26 Jun 2025 15:55 +0000
in reply to: https://benjojo.co.uk/u/benjojo/h/1gFkzQm2httXYv2G37
@benjojo trust me, you might want to search your photos by microwave or toaster, sometimes you want to check a serial or something that you took a picture in the past. I'm using immich, and the search function helped me out a few times, when I needed to check my car license plates for example, just searched for car lol
evey@chaos.social
replied 26 Jun 2025 15:55 +0000
in reply to: https://benjojo.co.uk/u/benjojo/h/1gFkzQm2httXYv2G37
@benjojo If they are trained on ImageNet the label set is somehwhay. Okay, but if you want anything more specific you want to cross train them against your own data.
Its mostly the limited avaiblity of good labeled datasets :/
alex@feed.yopp.me
replied 26 Jun 2025 16:20 +0000
in reply to: https://benjojo.co.uk/u/benjojo/h/1gFkzQm2httXYv2G37
@benjojo if you know labeled dataset of “normal things” it’s kinda-sorta easy to fine-tune existing YOLO model But yeah, it’s really weird that there is no single “label graph database” or something like that