I have looked at all the kernels and discussions and it seems that more people are doing segmentation and not detection. They are generating boxes from connected segments, but why aren't people just using an object detector like YOLO or RetinaNet? Is there not enough data? Is segmentation advantageous in this setting or is segmentation simply easier to do in a short time?
", "messages": [ { "id": "386456", "postDate": "09/12/2018 23:07:41", "content": "I have looked at all the kernels and discussions and it seems that more people are doing segmentation and not detection. They are generating boxes from connected segments, but why aren't people just using an object detector like YOLO or RetinaNet? Is there not enough data? Is segmentation advantageous in this setting or is segmentation simply easier to do in a short time?
", "rawMarkdown": "I have looked at all the kernels and discussions and it seems that more people are doing segmentation and not detection. They are generating boxes from connected segments, but why aren't people just using an object detector like YOLO or RetinaNet? Is there not enough data? Is segmentation advantageous in this setting or is segmentation simply easier to do in a short time?", "votes": null }, { "id": "386514", "postDate": "09/13/2018 03:46:28", "content": "Can't say, but i think we can follow both approch, segmentation and detection. I think segmentation is easy, and object detection is a little tricky.
", "rawMarkdown": "Can't say, but i think we can follow both approch, segmentation and detection. I think segmentation is easy, and object detection is a little tricky.", "votes": null }, { "id": "386560", "postDate": "09/13/2018 05:59:42", "content": "> I think segmentation is easy, and object detection is a little tricky.
\n\nYes, I think that's true.\nHowever it would be good if someone reports that an object detector like YOLO works good for this problem.
", "rawMarkdown": "> I think segmentation is easy, and object detection is a little tricky.\n\nYes, I think that's true.\nHowever it would be good if someone reports that an object detector like YOLO works good for this problem.", "votes": null }, { "id": "386593", "postDate": "09/13/2018 07:50:34", "content": "I think YOLO will work and its easy to be build and train.
", "rawMarkdown": "I think YOLO will work and its easy to be build and train.", "votes": null }, { "id": "387187", "postDate": "09/14/2018 14:00:21", "content": "The reason I'm trying to do segmentation is because of the annotation misalignment and label noise. I need to devise a strategy to learn the signal and not the noise.
", "rawMarkdown": "The reason I'm trying to do segmentation is because of the annotation misalignment and label noise. I need to devise a strategy to learn the signal and not the noise.", "votes": null }, { "id": "387232", "postDate": "09/14/2018 15:25:11", "content": "simply because most of the kernels with segmentation approach are forked! LOL.. if someone post a kernel using object detection with a good score, you will see fork object detection kernel.. :D
", "rawMarkdown": "simply because most of the kernels with segmentation approach are forked! LOL.. if someone post a kernel using object detection with a good score, you will see fork object detection kernel.. :D", "votes": null }, { "id": "387291", "postDate": "09/14/2018 17:20:10", "content": "I'm using retinanet for object detection. I think like people have said it's a matter of forking the highest scoring kernels and probably the fact that object detection for this task requires considerable compute time, for which kernels are rather limited.
", "rawMarkdown": "I'm using retinanet for object detection. I think like people have said it's a matter of forking the highest scoring kernels and probably the fact that object detection for this task requires considerable compute time, for which kernels are rather limited.", "votes": null }, { "id": "387633", "postDate": "09/15/2018 09:57:35", "content": "Did you get your score 0.134 by Retinanet?
", "rawMarkdown": "Did you get your score 0.134 by Retinanet?", "votes": null }, { "id": "387827", "postDate": "09/15/2018 18:16:21", "content": "yeah using following implementation https://github.com/fizyr/keras-retinanet
", "rawMarkdown": "yeah using following implementation https://github.com/fizyr/keras-retinanet", "votes": null }, { "id": "388741", "postDate": "09/17/2018 13:54:16", "content": "U-Net / Mask- RCNN should keep both sides happy
", "rawMarkdown": "U-Net / Mask- RCNN should keep both sides happy", "votes": null }, { "id": "388771", "postDate": "09/17/2018 15:03:05", "content": "Thats also a good idea, generate both bounding box and mask.
", "rawMarkdown": "Thats also a good idea, generate both bounding box and mask.", "votes": null } ], "comments": [ { "id": 386514, "author_name": "nikhilroxtomar", "author_url": "", "post_date": "09/13/2018 03:46:28", "content": "Can't say, but i think we can follow both approch, segmentation and detection. I think segmentation is easy, and object detection is a little tricky.
", "votes": null, "replies": [ { "id": 386560, "author_name": "sergeyzlobin", "author_url": "", "post_date": "09/13/2018 05:59:42", "content": "> I think segmentation is easy, and object detection is a little tricky.
\n\nYes, I think that's true.\nHowever it would be good if someone reports that an object detector like YOLO works good for this problem.
", "votes": null, "replies": [] }, { "id": 386593, "author_name": "nikhilroxtomar", "author_url": "", "post_date": "09/13/2018 07:50:34", "content": "I think YOLO will work and its easy to be build and train.
", "votes": null, "replies": [] } ] }, { "id": 387187, "author_name": "puremath86", "author_url": "", "post_date": "09/14/2018 14:00:21", "content": "The reason I'm trying to do segmentation is because of the annotation misalignment and label noise. I need to devise a strategy to learn the signal and not the noise.
", "votes": null, "replies": [] }, { "id": 387232, "author_name": "", "author_url": "", "post_date": "09/14/2018 15:25:11", "content": "simply because most of the kernels with segmentation approach are forked! LOL.. if someone post a kernel using object detection with a good score, you will see fork object detection kernel.. :D
", "votes": null, "replies": [] }, { "id": 387291, "author_name": "taindow", "author_url": "", "post_date": "09/14/2018 17:20:10", "content": "I'm using retinanet for object detection. I think like people have said it's a matter of forking the highest scoring kernels and probably the fact that object detection for this task requires considerable compute time, for which kernels are rather limited.
", "votes": null, "replies": [ { "id": 387633, "author_name": "sergeyzlobin", "author_url": "", "post_date": "09/15/2018 09:57:35", "content": "Did you get your score 0.134 by Retinanet?
", "votes": null, "replies": [] }, { "id": 387827, "author_name": "taindow", "author_url": "", "post_date": "09/15/2018 18:16:21", "content": "yeah using following implementation https://github.com/fizyr/keras-retinanet
", "votes": null, "replies": [] } ] }, { "id": 388741, "author_name": "bcsunil", "author_url": "", "post_date": "09/17/2018 13:54:16", "content": "U-Net / Mask- RCNN should keep both sides happy
", "votes": null, "replies": [ { "id": 388771, "author_name": "nikhilroxtomar", "author_url": "", "post_date": "09/17/2018 15:03:05", "content": "Thats also a good idea, generate both bounding box and mask.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "386456": "I have looked at all the kernels and discussions and it seems that more people are doing segmentation and not detection. They are generating boxes from connected segments, but why aren't people just using an object detector like YOLO or RetinaNet? Is there not enough data? Is segmentation advantageous in this setting or is segmentation simply easier to do in a short time?", "386514": "Can't say, but i think we can follow both approch, segmentation and detection. I think segmentation is easy, and object detection is a little tricky.", "386560": "> I think segmentation is easy, and object detection is a little tricky.\n\nYes, I think that's true.\nHowever it would be good if someone reports that an object detector like YOLO works good for this problem.", "386593": "I think YOLO will work and its easy to be build and train.", "387187": "The reason I'm trying to do segmentation is because of the annotation misalignment and label noise. I need to devise a strategy to learn the signal and not the noise.", "387232": "simply because most of the kernels with segmentation approach are forked! LOL.. if someone post a kernel using object detection with a good score, you will see fork object detection kernel.. :D", "387291": "I'm using retinanet for object detection. I think like people have said it's a matter of forking the highest scoring kernels and probably the fact that object detection for this task requires considerable compute time, for which kernels are rather limited.", "387633": "Did you get your score 0.134 by Retinanet?", "387827": "yeah using following implementation https://github.com/fizyr/keras-retinanet", "388741": "U-Net / Mask- RCNN should keep both sides happy", "388771": "Thats also a good idea, generate both bounding box and mask." }, "source": "meta" }