The following is a collection of high-perofmring object detection models with permissive licenses (e.g. Apache 2.0, MIT).
Why collect these?
After diving into the world of object detection, you’ll probably run into YOLO (You Only Look Once) models.
These models perform very well, however, they often come with the AGPL-3.0 license.
I’m not a lawyer but when I read this license, it seems that using code (including models) from repos containing this license results in you having to publish your own code under AGPL-3.0 as well.
This means if you’re a startup or a business using a AGPL-3.0 model and your app uses that model in a certain way, you must publish the app code as open-source under AGPL-3.0 too.
TL;DR
If you want to keep your source code closed source or just generally want a more permissive license, use Apache 2.0 or MIT code/models.
Again, I’m not a lawyer but as a developer, I just prefer using things I know aren’t gonna come back to bite me.
Apache 2.0 or MIT licensed object detection models
I’ve personally vetted the following to perform on par or better than almost any YOLO variant for real-time object detection.
| Name | Code/Models | Paper | License | Notes |
|---|---|---|---|---|
| RF-DETR (Roboflow DETR) | GitHub | - | Apache 2.0 | |
| RT-DETRv2 (Real-time DETR) | GitHub, Hugging Face | arXiv | Apache 2.0 | |
| D-FINE (Fine‑grained Distribution Refinement) | GitHub, Hugging Face | arXiv | Apache 2.0 |
Notes:
- DETR = DEctection TRansformer (using the Transformer model architecture for object detection)
- Real-time object detection = models capable of identifying objects in image(s) at 24FPS or higher, once they reach this threshold, these models are capable of identifying objects in video feeds, hence the “real-time” description