DINOv3-B teacher model

[SebastianJanampa]

Hi,

Thanks for the great work. Are you planning on realizing the DINOv3-B teacher model weights?

[capsule2077]

Hi, thank you for your interest! Are you referring to the weights of ECTeacher-B, the teacher model used in our distillation process?

[SebastianJanampa]

Hi,
Yes, I am referring to ECTeacher-B. Also, why did you use ECTeacher-S instead of ECTeacher-M for the smallest model?

[capsule2077]

1. Here is the link to download ECTeacher-B. By the way, it shares the same structure as DINOv3-B but was pre-trained on COCO.

2. Just to clarify, our naming follows DINOv3 (S/B/L) because they share the same network structures and only differ in weights, so there is no M version.

[SebastianJanampa]

Thanks for sharing the weights.

I apologize for the misunderstanding. I accidentally mispelled. What I meant to say was "Why did you use ECTeacher-S instead of ECTeacher-B for the smallest model?"

[capsule2077]

We've observed that an excessive capacity gap between teacher and student leads to performance drops. Specifically, ECTeacher-B performs worse than ECTeacher-S when distilling ecdet-s. You can find the supporting data in Table 5 of the report.

[HebeiFast]

Why is there such a large gap between the Teacher model and the Student model, yet it's still possible to distill a model that outperforms the teacher? Could the authors provide some explanation for this? Also, would it be possible to open-source the code for the detection part?

[chcchc97]

@capsule2077

1. Here is the link to download ECTeacher-B. By the way, it shares the same structure as DINOv3-B but was pre-trained on COCO.
2. Just to clarify, our naming follows DINOv3 (S/B/L) because they share the same network structures and only differ in weights, so there is no M version.

@capsule2077 Thanks for sharing! I was wondering if you could provide the training logs for ECTeacher-B? I'm looking to reproduce the results. Thank you very much!

[SebastianJanampa]

Hi @capsule2077 ,

Could you also provide the results (mAP, FLOPs, # of Params, latency) on the COCO datasets for ECTeacher-S/B/L?