Detectron2: Any Plan to open source PointRend?

[txytju]

[ppwwyyxx]

PointRend for instance segmentation will be available in a month.

[startupgurukul]

If we notice in the gif, Densepose isnt very accurate at the edges, hoping that would also improve with the PointRend

[zsef123]

PointRend-PyTorch

Now I'm working for Semantic Segmentation.
But it's not fully implementation and not tested yet.

[corporate-slave]

PointRend-PyTorch

Now I'm working for Semantic Segmentation.
But it's not fully implementation and not tested yet.

Can I download this and start testing it out? Or the library isnt working at all?

[zsef123]

@corporate-slave

You can running only train networks.
But I checked only a few steps run(loss decreases well)
And I'm not implement inference phase.

[tomer-avrahami]

Hi @ppwwyyxx,

PointRend paper describes different point selection techniques for training and inference. In addition, different parameters were used for different tasks.
Can you please share how much effect it had on your final results for instance segmentation? Did you see a major difference in your final results or stability improvements with different techniques?
Do you think that better technique could get better results?

Thanks,
Tomer

[jay2002]

Hi @ppwwyyxx
Is the PointRend available?

[xmyqsh]

@zsef123
You have misunderstood the point sampling process.
Maybe I can help you.

[zsef123]

@xmyqsh
issue or PR are welcome anytime

[xmyqsh]

@zsef123
I'm implementing PointRend for instance segmentation.
I will help you after that.

[bluesky314]

@zsef123 Is the implementation complete?

[rbgirshick]

We've released code yesterday for PointRend instance segmentation under detectron2/projects here: https://github.com/facebookresearch/detectron2/tree/master/projects/PointRend.

The PointRend semantic segmentation code is nearly ready for release and will be available soon.

[ppwwyyxx]

FYI we just released a colab notebook that visualizes the internal steps of PointRend: https://colab.research.google.com/drive/1isGPL5h5_cKoPPhVL9XhMokRtHDvmMVL

[bluesky314]

Awesome! I will go through this.

On Sat, Feb 8, 2020 at 5:55 AM Yuxin Wu [e-mail] wrote:

FYI we just released a colab notebook that visualizes the internal steps
of PointRend:
https://colab.research.google.com/drive/1isGPL5h5_cKoPPhVL9XhMokRtHDvmMVL

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
https://github.com/facebookresearch/detectron2/issues/581?email_source=notifications&email_token=AIOBDNBHE73OSW4CVBDP3K3RBX3X7A5CNFSM4J7DBZY2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOELFCJHQ#issuecomment-583672990,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AIOBDNHCBVJGK5YARY6HZWDRBX3X7ANCNFSM4J7DBZYQ
.