search

yoxu515 / aot-benchmark

An efficient modular implementation of Associating Objects with Transformers for Video Object Segmentation in PyTorch
BSD 3-Clause "New" or "Revised" License
600 stars 108 forks source link

about DMAOT details? #62

Open zhanghongyong123456 opened 1 year ago

zhanghongyong123456 commented 1 year ago

I saw DMAOT related news some time ago, is it not planned to be released?

bhack commented 11 months ago

Any news on this?

z-x-yang commented 11 months ago

Thanks for your interest. We plan to release the implementation of DM-AOT in the next several weeks.

bhack commented 10 months ago

Have DMAOT also used the multiscale MS-DeAOT?

bhack commented 10 months ago

@z-x-yang @yoxu515 Can you disclose something more on this? Is there any technical report other then the ICCV slides?

yamy-cheng commented 10 months ago

@z-x-yang @yoxu515 Can you disclose something more on this? Is there any technical report other then the ICCV slides?

Hi, the multiscale strategy has not been utilized in DMAOT, and the code will be released in the next few days.

bhack commented 10 months ago

There was any specific motivation why you have not included the multiscale strategy in DMAOT?

yamy-cheng commented 10 months ago

The multiscale strategy requires higher GPU utilization, but unfortunately, we lack the necessary resources to train the SwinB-MS-DeAOT-L. Additionally, the R50-MS-DeAOT-L model does not outperform the SwinB-DeAOT-L model. Therefore, we have decided to use SwinB-DeAOT-L as the base model. The code of DMAOT is now open and can be found here.

zhanghongyong123456 commented 10 months ago

The multiscale strategy requires higher GPU utilization, but unfortunately, we lack the necessary resources to train the SwinB-MS-DeAOT-L. Additionally, the R50-MS-DeAOT-L model does not outperform the SwinB-DeAOT-L model. Therefore, we have decided to use SwinB-DeAOT-L as the base model. The code of DMAOT is now open and can be found here.

请问有论文发布吗，没有找到论文链接