Information about dataset to test models' robustness

[SebastianJanampa]

Hi Roboflow team,

Thanks for the continuous support of RF-DETR, and congratulations on RF-DETR-Pose.

In the RF-DETR paper, you mention using RF100-VL to assess how well models generalize to other datasets. I want to ask if there's any other dataset to test the robustness of COCO-trained models, since you fine-tuned all the models on the RF100-VL dataset.

Best,
Sebastian

[isaacrob]

Honestly there aren't any datasets that are in common use. We hope people adopt RF100-VL

[SebastianJanampa]

I see. Thanks for the answer. I tried to train DEIMv2 on the RF100-VL, but training the small version took a long time, making it impossible to train the other variants. This is a point in favor of RF-DETR, since you only need to train once and can use a wide range of models.

[isaacrob]

Training on all of RF100-VL should take about as long as training on COCO. Total dataset sizes are similar.