search

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.28k stars 630 forks source link

Do you have a plan for instance segmentation? #94

Closed stigma0617 closed 5 years ago

stigma0617 commented 5 years ago

Hi,

In your paper,

"The proposed detector can be immediately extended to solve other vision tasks with minimal modification, including instance segmentation and key-point detection."

is described.

Do you have a plan to extend for instance segmentation such as Mask R-CNN ?

tianzhi0549 commented 5 years ago

@stigma0617 We are working on instance segmentation now.

stigma0617 commented 5 years ago

Cool !! Your work is very awesome!! good luck !

tianzhi0549 commented 5 years ago

@stigma0617 Thank you.

stigma0617 commented 5 years ago

@tianzhi0549 Congratulation for acception in ICCV'19.

When do you expect the instance segmentation to be completed?

tianzhi0549 commented 5 years ago

@stigma0617 Thank you for your congratulations. I can't give you a guarantee. We are trying some new ideas for instance segmentation (not only using maskrcnn), so I am not sure the time of completion.

stigma0617 commented 5 years ago

@tianzhi0549 ,

Oh I see, I'm developing my new VoVNet backbone in your FCOS. So I just wondered instance segmentation task.

my naive result(VoV57-FCOS) is that

I will also upload the VoVNet-FCOS as soon as possible.

Thank you for your good work again :)

Good luck 👍

tianzhi0549 commented 5 years ago

@stigma0617 Great work! Thank you very much for sharing your results and code.

heng2j commented 5 years ago

Hi @stigma0617, your results with VoVNet-FCOS, I’m wonder in your repository, will you also share about your insights of how to retain the model with your own dataset? And how’s the configuration should be for VoVNet?

stigma0617 commented 5 years ago

@heng2j I trained VoVNet-FCOS on COCO dataset.

The configuration of VoVNet-FCOS is same as VoVNet-Detectron.

heng2j commented 5 years ago

Hi @stigma0617.

Thank you for your replying! Like what you replied on the I am currently training a FCOS_PLUS model with my own dataset and used VoVNet57 as the backbone.

I really want to see how well it perform on small objects.

heng2j commented 5 years ago

So after 82000 iterations, the AP are still very low. And the loss decays so slow as well.

stigma0617 commented 5 years ago

@heng2j, I'll upload VoVNet-FCOS soon. When I upload the CODE, I'll notify you.

heng2j commented 5 years ago

Hi @stigma0617, thank you for the updates! Looking forward! At the meantime, I will stick with RestNet first for the baseline.

stigma0617 commented 5 years ago

@heng2j , @tianzhi0549

I upload the VoVNet-FCOS project.

heng2j commented 5 years ago

Hi @stigma0617 , thank you so much for letting us know! Greatly Appreciated! I will take a look and post issue on your repo if I encountered any issue. Cheers!

tianzhi0549 commented 5 years ago

@stigma0617 Thank you very much! I will add it into our README to let more people know this.

stigma0617 commented 5 years ago

@tianzhi0549 Thanks :)