Has anyone experimented with using a CRF for prediction modification at high resolution? I'm still working with 256x256 size images, but it's made predictions worse with every architecture I have tried it with-- looks like a dead end.
", "messages": [ { "id": "221224", "postDate": "09/14/2017 14:52:42", "content": "Has anyone experimented with using a CRF for prediction modification at high resolution? I'm still working with 256x256 size images, but it's made predictions worse with every architecture I have tried it with-- looks like a dead end.
", "rawMarkdown": "Has anyone experimented with using a CRF for prediction modification at high resolution? I'm still working with 256x256 size images, but it's made predictions worse with every architecture I have tried it with-- looks like a dead end.", "votes": null }, { "id": "221238", "postDate": "09/14/2017 15:55:11", "content": "I was thinking about trying them, but I have no experience with graphical models. How do you implement a CRF? Is there some dedicated package to conduct inference?
", "rawMarkdown": "I was thinking about trying them, but I have no experience with graphical models. How do you implement a CRF? Is there some dedicated package to conduct inference?", "votes": null }, { "id": "221251", "postDate": "09/14/2017 16:35:43", "content": "There are a couple of opensource implementations. Pydensecrf is a good starter package, but running it vanilla will not give you good results on this problem (-.002 to -.0025 per image). Even with modifications I could not get an improvement out of it.
", "rawMarkdown": "There are a couple of opensource implementations. Pydensecrf is a good starter package, but running it vanilla will not give you good results on this problem (-.002 to -.0025 per image). Even with modifications I could not get an improvement out of it.", "votes": null }, { "id": "221341", "postDate": "09/14/2017 21:13:29", "content": "I was getting improvement with Pydensecrf with some modification only for low score images. I.e. dice score 0.990 -> 0.992. But for high score images the same settings was giving lower score, i.e 0.997 -> 0.996. The biggest prob though (besides determine good and bad images on test), is that Pydensecrf is really slow and it's hard to make experiments.
", "rawMarkdown": "I was getting improvement with Pydensecrf with some modification only for low score images. I.e. dice score 0.990 -> 0.992. But for high score images the same settings was giving lower score, i.e 0.997 -> 0.996. The biggest prob though (besides determine good and bad images on test), is that Pydensecrf is really slow and it's hard to make experiments.", "votes": null }, { "id": "221509", "postDate": "09/15/2017 14:16:28", "content": "Is it OK to choose manually which one applied CRF processing?
", "rawMarkdown": "Is it OK to choose manually which one applied CRF processing?", "votes": null }, { "id": "221673", "postDate": "09/16/2017 02:16:08", "content": "I had tried dense crf in deeplab v2. However, I get a worse result than end-to-end predition.
", "rawMarkdown": "I had tried dense crf in deeplab v2. However, I get a worse result than end-to-end predition.", "votes": null }, { "id": "282591", "postDate": "02/14/2018 08:02:24", "content": "Can anyone share how to use Pydensecrf? Thanks in advance.
", "rawMarkdown": "Can anyone share how to use Pydensecrf? Thanks in advance.", "votes": null } ], "comments": [ { "id": 221238, "author_name": "stijndecubber", "author_url": "", "post_date": "09/14/2017 15:55:11", "content": "I was thinking about trying them, but I have no experience with graphical models. How do you implement a CRF? Is there some dedicated package to conduct inference?
", "votes": null, "replies": [ { "id": 221251, "author_name": "rhgrossm", "author_url": "", "post_date": "09/14/2017 16:35:43", "content": "There are a couple of opensource implementations. Pydensecrf is a good starter package, but running it vanilla will not give you good results on this problem (-.002 to -.0025 per image). Even with modifications I could not get an improvement out of it.
", "votes": null, "replies": [] }, { "id": 221341, "author_name": "heyt0ny", "author_url": "", "post_date": "09/14/2017 21:13:29", "content": "I was getting improvement with Pydensecrf with some modification only for low score images. I.e. dice score 0.990 -> 0.992. But for high score images the same settings was giving lower score, i.e 0.997 -> 0.996. The biggest prob though (besides determine good and bad images on test), is that Pydensecrf is really slow and it's hard to make experiments.
", "votes": null, "replies": [] }, { "id": 221509, "author_name": "lyakaap", "author_url": "", "post_date": "09/15/2017 14:16:28", "content": "Is it OK to choose manually which one applied CRF processing?
", "votes": null, "replies": [] } ] }, { "id": 221673, "author_name": "kinpzz", "author_url": "", "post_date": "09/16/2017 02:16:08", "content": "I had tried dense crf in deeplab v2. However, I get a worse result than end-to-end predition.
", "votes": null, "replies": [] }, { "id": 282591, "author_name": "jamesluo1991", "author_url": "", "post_date": "02/14/2018 08:02:24", "content": "Can anyone share how to use Pydensecrf? Thanks in advance.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "221224": "Has anyone experimented with using a CRF for prediction modification at high resolution? I'm still working with 256x256 size images, but it's made predictions worse with every architecture I have tried it with-- looks like a dead end.", "221238": "I was thinking about trying them, but I have no experience with graphical models. How do you implement a CRF? Is there some dedicated package to conduct inference?", "221251": "There are a couple of opensource implementations. Pydensecrf is a good starter package, but running it vanilla will not give you good results on this problem (-.002 to -.0025 per image). Even with modifications I could not get an improvement out of it.", "221341": "I was getting improvement with Pydensecrf with some modification only for low score images. I.e. dice score 0.990 -> 0.992. But for high score images the same settings was giving lower score, i.e 0.997 -> 0.996. The biggest prob though (besides determine good and bad images on test), is that Pydensecrf is really slow and it's hard to make experiments.", "221509": "Is it OK to choose manually which one applied CRF processing?", "221673": "I had tried dense crf in deeplab v2. However, I get a worse result than end-to-end predition.", "282591": "Can anyone share how to use Pydensecrf? Thanks in advance." }, "source": "meta" }