search

IDEA-Research / DINO

[ICLR 2023] Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection"
Apache License 2.0
2.16k stars 234 forks source link

The performance on Object365 #109

Closed chensnathan closed 1 year ago

chensnathan commented 2 years ago

Hi,

Have the pre-trained DINO-SwinL been evaluated on the Object365 validation set? It would be much appreciated if it is possible to share the performance on the object365 dataset.

Chen-Song commented 2 years ago

@SlongLiu @FengLi-ust 同楼上，我也想知道DINO-SwinL在Object365验证集表现。另外请问可以放出DINO-SwinL的Object365预训练模型吗？

FengLi-ust commented 1 year ago

The performance of O365 val depends on the val images you choose. In our val set, the performance is around 45-47 AP. We will consider releasing the checkpoint of this large model after this paper gets peer-reviewed.

SlongLiu commented 1 year ago

We use the first 5000 images in O365 as our val set, as mentioned in our paper.

chensnathan commented 1 year ago

Got it. Thanks for sharing the details and results. It helps a lot.