roboflow
RF-DETR shows within domain generalizability
For a research project, RF-DETR-L was trained on a large train dataset and subsequently evaluated on its corresponding benchmark test dataset, the performance ($mAP_w$) was similar to the SOTA model in this specific domain.
We created a new benchmark dataset, containing the same objects, but in a different setting (e.g., viewpoint, lightning) and with different types of the same objects compared to the aforementioned dataset. All models that we evaluated were only trained on the aforementioned dataset.
Next to exhibiting across domain adaptability of RF-DETR, as you have shown by benchmarking it on RF100-VL. RF-DETR-L also seems to be good at within domain generalizability, as it outperformed the SOTA model (+8 $mAP_w$), without additional fine-tuning on part of the new benchmark dataset. This within domain generalizability might be due to the DINOv2 based backbone?
Really nice to see, hopefully this research will be published, as this current description is quite vague.
hi @DatSplit! glad to hear it's working well for you! I'd love to know about which domain you're applying it to :) the model was designed to generalize nicely which seems to be what you're seeing, great for us to have that real world feedback!
the domain generalizability is likely due to several factors, one of which is the DINOv2 based backbone, and the others we will discuss in the upcoming paper! we haven't released it yet because we're still building, with lots of improvements coming in the (hopefully near) future. we're a VERY small team working on this so it's taking us a while to finish, please have patience with us! :)
Hi @isaacrob-roboflow!
We are using RF-DETR in the fashion apparel (clothes and accessories) domain. Great work that you have done so far, especially as a very small team!
Do you maybe have a ballpark figure for the release date of the paper on RF-DETR?
unfortunately no :/ we're planning on doing several more feature releases before the paper release, and that progress is prone to interruption by needs from other parts of the company. very chaotic being a tiny team within a small startup! :)
@DatSplit can you say a bit about why you're eager for the paper? Might help justify our working on it sooner :)
Hi @isaacrob-roboflow,
We are curious why there might be "within domain generalization" compared to other models. We will examine this by doing some visual empirical experiments. E.g., we expect RF-DETR-L to be better at detecting smart watches compared to other existing models.
I guess specifically, is there any actionable that is being blocked by our not having a paper out? or is it just interest :)
Good morning @isaacrob-roboflow ,
It is just out of interest! Again, thank you and your colleagues for this awesome model!
