Move solo to mmdetection2

[JarintotionDin]

How about move solo to mmdetection2

[WXinlong]

@JarintotionDin The SOLO implementation in mmdet v2 is at here.

[JarintotionDin]

@WXinlong Thanks so much !

[kemaloksuz]

Do you have mmdet v2 implementation of solov2 as well?

[stefOns]

@kemaloksuz An implementation of SOLOv2 with mmdet 2.3 can be found here. We extended the SOLO branch from here.

[kemaloksuz]

Many thanks! I will check it out.

[danielcrane]

@kemaloksuz An implementation of SOLOv2 with mmdet 2.3 can be found here. We extended the SOLO branch from here.

Thanks for sharing this, out of curiosity are there any benefits to using mmdet 2.3? Or is it just helpful for compatibility reasons?

[kemaloksuz]

@kemaloksuz An implementation of SOLOv2 with mmdet 2.3 can be found here. We extended the SOLO branch from here.

Thanks for sharing this, out of curiosity are there any benefits to using mmdet 2.3? Or is it just helpful for compatibility reasons?

For me, from SOLOv2 point of view, it was the compatibility. We have released a new repository for our ICCV paper recently (https://github.com/kemaloksuz/RankSortLoss), in which we could extend all models with our method in mmdet 2.8 except SOLOv2 since at that time we could not find an mmdet 2.X implementation of SOLOv2 and extended a separate repo only for SOLOv2 on mmdet 1.0.0 (https://github.com/WXinlong/SOLO). When I have time, I will try to add this mmdet 2.3 implementation to our repo, which I expect to be easier owing to better compatibility.

If you are asking the benefits of different versions of mmdetection in general, then I would recommend you checking out this changelog: https://github.com/open-mmlab/mmdetection/blob/master/docs/changelog.md

[stefOns]

@kemaloksuz An implementation of SOLOv2 with mmdet 2.3 can be found here. We extended the SOLO branch from here.

Thanks for sharing this, out of curiosity are there any benefits to using mmdet 2.3? Or is it just helpful for compatibility reasons?

In the first place it was necessary for compatibility reasons, since we are using new technologies (e.g torchserve) and we run into problems with several dependencies to older versions (python, pytorch, mmcv,...). mmdet 2.x is said to be faster but I dind't compare the runtime. From the instance segmentation results I didn't realize that improvement...

[danielcrane]

@kemaloksuz An implementation of SOLOv2 with mmdet 2.3 can be found here. We extended the SOLO branch from here.

Thanks for sharing this, out of curiosity are there any benefits to using mmdet 2.3? Or is it just helpful for compatibility reasons?

In the first place it was necessary for compatibility reasons, since we are using new technologies (e.g torchserve) and we run into problems with several dependencies to older versions (python, pytorch, mmcv,...). mmdet 2.x is said to be faster but I dind't compare the runtime. From the instance segmentation results I didn't realize that improvement...

Thanks so much, I appreciate your response!