After some time after hearing rumors about Dan Morris working on new MegaDetector models, they’re finally here! We were contacted by Dan two weeks ago who offered to us to test out the new models and give some feedback. Frankly, that was in the middle of my vacation, so my time to test it out back then was limited. Since then, there has been a silent release of the new models, and I belive it needs some more attention!
The new models are released under the name “MDv1000” and here is what to expect.
MDv5a, the predecessor still works excellent and maybe even better in some cases, but MDv1000-redwood has taken the leaderboard of best naAP across a wide range of wildlife trail camera images.
The issues that has been specifically targeted with the new models, citing Dan are:
"I consider these models to be an incremental change to MDv5, i.e., they are still just animal/person/vehicle and they don't add any fundamentally new capabilities, rather, I specifically targeted (a) this issue ( https://lnkd.in/dVkQ853s) that was driving me crazy by sometimes putting random boxes in the sky and missing animals and (b) improved performance on reptiles, as well as a bunch of other issues in the public portion of the MDv5 training data that I found once I started looking (particularly a lot of false negatives in the labels on small birds and rodents). “
And while I personally have not been able to test all of the claims, I for sure see an increase of detection of reptiles.
There are other different models, which works different, in different use cases.
MDv1000-redwood (YOLOv5x6, 1280px, naAP 1.01)
MDv1000-cedar (YOLOv9c, 640px, naAP 0.99)
MDv1000-larch (YOLOv11L, 640px, naAP 0.97)
MDv1000-sorrel (YOLOv11s, 960px, naAP 0.97)
MDv1000-spruce (YOLOv5s, 640px, naAP 0.86)
For comparison:
MDv5a (YOLOv5x6, 1280px, naAP 1.00 by definition)
MDv5b (YOLOv5x6, 1280px, naAP 0.99)
So, I briefly already introduced MDv1000-redwood which in most cases will be the new "go-to", I have only personally tested MDv1000-cedar, which is significantly faster, with a smaller YOLOv9c model. While testing it on global images from trail cameras, it works wonderfully. I and Dan tested it on a bit special dataset where the MDv1000-cedar seems to find quite some challenging detections which the MDv5 didn’t but fell short in other images.
If you want a more comprehensive read about the models, please check out the release-notes here: https://lnkd.in/dJtr3aQ4
Images taken from the release notes as they illustrates some of the issues and fixes, Image credit: UNSW Predators dataset for the reptile image and Image credit Snapshot Serengeti dataset for the floating BBox images.
