![\"\"](\"https://i.ibb.co/yQJVMR4/Selection-060.png\")
\n
![\"\"](\"https://i.ibb.co/tKys61P/Selection-074.png\")
\n
![\"\"](\"https://i.ibb.co/xffJpSP/Selection-062.png\")
\n
![\"\"](\"https://i.ibb.co/YXNXmFM/Selection-063.png\")
for those who have tried segmentation, you may want to share your experiences here.
\nI personally feel that segmentation is more difficult. This is because a line long. You cannot differentiate the class label if you are at \"a certain part of the line\". Some parts are easier to differentiate (e.g. the endpoint, loop, etc), but some are almost non-discriminative.
\nyou need to have a very large context or some way to capture the relationships of all parts of line (e.g. link all points to all points ). In theory, the transformer can work but i haven't got time to play with it yet.
\n
\n
\n
\n
it is easier just to segment important parts of the line, instead of the whole line. But identifying the discriminative parts itself is a problem. (e.g. consider top-k pixel labels for the positive case within group of ett,ngt,cvc for loss and inference)
\none can refer to DETR set loss that is used to remove NMS in object detection.Here. we treat the ground truth line mask as a set of possible +ve pixel label. we just need at least top-k +ve pixel label during training, (and inference)
", "messages": [ { "id": "1228301", "postDate": "03/06/2021 08:53:40", "content": "for those who have tried segmentation, you may want to share your experiences here.
\nI personally feel that segmentation is more difficult. This is because a line long. You cannot differentiate the class label if you are at \"a certain part of the line\". Some parts are easier to differentiate (e.g. the endpoint, loop, etc), but some are almost non-discriminative.
\nyou need to have a very large context or some way to capture the relationships of all parts of line (e.g. link all points to all points ). In theory, the transformer can work but i haven't got time to play with it yet.
\n
\n
\n
\n
it is easier just to segment important parts of the line, instead of the whole line. But identifying the discriminative parts itself is a problem. (e.g. consider top-k pixel labels for the positive case within group of ett,ngt,cvc for loss and inference)
\none can refer to DETR set loss that is used to remove NMS in object detection.Here. we treat the ground truth line mask as a set of possible +ve pixel label. we just need at least top-k +ve pixel label during training, (and inference)
", "rawMarkdown": "for those who have tried segmentation, you may want to share your experiences here.\n\nI personally feel that segmentation is more difficult. This is because a line long. You cannot differentiate the class label if you are at \"a certain part of the line\". Some parts are easier to differentiate (e.g. the endpoint, loop, etc), but some are almost non-discriminative.\n\nyou need to have a very large context or some way to capture the relationships of all parts of line (e.g. link all points to all points ). In theory, the transformer can work but i haven't got time to play with it yet.\n\n\n\n\n\n\nit is easier just to segment important parts of the line, instead of the whole line. But identifying the discriminative parts itself is a problem. (e.g. consider top-k pixel labels for the positive case within group of ett,ngt,cvc for loss and inference)\n\none can refer to DETR set loss that is used to remove NMS in object detection.Here. we treat the ground truth line mask as a set of possible +ve pixel label. we just need at least top-k +ve pixel label during training, (and inference)", "votes": null }, { "id": "1228644", "postDate": "03/06/2021 16:05:45", "content": "In my experience the problem is like what you were saying, some parts are easy to segment, but they do not provide any value, like the long run of the line. The hard part to accurately segment is the final endpoint and the model shows a bit of uncertainty about that and that is the most important piece of the puzzle according to some of the experts.
\nWhat I found is that the segmentation mask for something like ETT will have a solid, but fading line that goes from top to bottom as it loses confidence towards the end of the tube placement.
", "rawMarkdown": "In my experience the problem is like what you were saying, some parts are easy to segment, but they do not provide any value, like the long run of the line. The hard part to accurately segment is the final endpoint and the model shows a bit of uncertainty about that and that is the most important piece of the puzzle according to some of the experts. \n\nWhat I found is that the segmentation mask for something like ETT will have a solid, but fading line that goes from top to bottom as it loses confidence towards the end of the tube placement.", "votes": null }, { "id": "1228707", "postDate": "03/06/2021 17:01:24", "content": "one can pseudo label the mask annotation by combining mask prediction and given truth image label
\n![\"\"](\"https://i.ibb.co/yXWP2sJ/Selection-089.png\")
i have been thinking about instance segmentation as oppose to pixel segmentation. I think it can be done as follows:
\n![\"\"](\"https://i.ibb.co/kyMZWgJ/Selection-116.png\")
@ryches ,
\nthe same by human observer :) One loses confidence towards the tip of the ETT…
\nMaybe a little bit of scatter reduction could help by that?
In my experience the problem is like what you were saying, some parts are easy to segment, but they do not provide any value, like the long run of the line. The hard part to accurately segment is the final endpoint and the model shows a bit of uncertainty about that and that is the most important piece of the puzzle according to some of the experts.
\nWhat I found is that the segmentation mask for something like ETT will have a solid, but fading line that goes from top to bottom as it loses confidence towards the end of the tube placement.
", "votes": null, "replies": [ { "id": 1232052, "author_name": "sandorkonya", "author_url": "", "post_date": "03/09/2021 13:05:00", "content": "@ryches ,
\nthe same by human observer :) One loses confidence towards the tip of the ETT…
\nMaybe a little bit of scatter reduction could help by that?
one can pseudo label the mask annotation by combining mask prediction and given truth image label
\n
i have been thinking about instance segmentation as oppose to pixel segmentation. I think it can be done as follows:
\n