Hey, how have you guys been handling image segmentation? I'm looking to use a sliding windows + CNN approach, but I want to have a plan for the full pipeline before implementing it. When we're using sliding windows for classification, and we've got a section of overlapping windows that all were classified as e.g. baby sea lions, how can we know how many baby sea lions are in that area?
\n\nI thought about dividing the size of that area by the size of a baby sea lion, but that seems imprecise. I also thought about applying an image segmentation algorithm, but it seems like those generally require image masks in the training data.
\n\nAny chance someone could point me in the right direction for this part of the challenge?
", "messages": [ { "id": "178763", "postDate": "04/28/2017 20:37:02", "content": "Hey, how have you guys been handling image segmentation? I'm looking to use a sliding windows + CNN approach, but I want to have a plan for the full pipeline before implementing it. When we're using sliding windows for classification, and we've got a section of overlapping windows that all were classified as e.g. baby sea lions, how can we know how many baby sea lions are in that area?
\n\nI thought about dividing the size of that area by the size of a baby sea lion, but that seems imprecise. I also thought about applying an image segmentation algorithm, but it seems like those generally require image masks in the training data.
\n\nAny chance someone could point me in the right direction for this part of the challenge?
", "rawMarkdown": "Hey, how have you guys been handling image segmentation? I'm looking to use a sliding windows + CNN approach, but I want to have a plan for the full pipeline before implementing it. When we're using sliding windows for classification, and we've got a section of overlapping windows that all were classified as e.g. baby sea lions, how can we know how many baby sea lions are in that area? \n\nI thought about dividing the size of that area by the size of a baby sea lion, but that seems imprecise. I also thought about applying an image segmentation algorithm, but it seems like those generally require image masks in the training data. \n\nAny chance someone could point me in the right direction for this part of the challenge?", "votes": null }, { "id": "178768", "postDate": "04/28/2017 20:43:42", "content": "\n\n\nhow can we know how many baby sea lions are in that area?
\n
One approach would be to sum the number of locations that give baby sea lion probability above some threshold, do that for several thresholds (e.g. 0.05, 0.25, 0.5), and use that as features for a regression algorithm (e.g. xgboost or lasso) that you train on your validation set.
", "rawMarkdown": "> how can we know how many baby sea lions are in that area?\n\nOne approach would be to sum the number of locations that give baby sea lion probability above some threshold, do that for several thresholds (e.g. 0.05, 0.25, 0.5), and use that as features for a regression algorithm (e.g. xgboost or lasso) that you train on your validation set.", "votes": null }, { "id": "178827", "postDate": "04/28/2017 23:32:18", "content": "Little hint: every sea lion female gives burth to one baby. Any sea lion pups must be always near its mother (female).
", "rawMarkdown": "Little hint: every sea lion female gives burth to one baby. Any sea lion pups must be always near its mother (female).", "votes": null }, { "id": "178970", "postDate": "04/29/2017 13:37:27", "content": "Nice, that sounds like it could work. Thanks for your help!
", "rawMarkdown": "Nice, that sounds like it could work. Thanks for your help!", "votes": null } ], "comments": [ { "id": 178768, "author_name": "lopuhin", "author_url": "", "post_date": "04/28/2017 20:43:42", "content": "\n\n\nhow can we know how many baby sea lions are in that area?
\n
One approach would be to sum the number of locations that give baby sea lion probability above some threshold, do that for several thresholds (e.g. 0.05, 0.25, 0.5), and use that as features for a regression algorithm (e.g. xgboost or lasso) that you train on your validation set.
", "votes": null, "replies": [ { "id": 178970, "author_name": "jamesthornton", "author_url": "", "post_date": "04/29/2017 13:37:27", "content": "Nice, that sounds like it could work. Thanks for your help!
", "votes": null, "replies": [] } ] }, { "id": 178827, "author_name": "sparamonov", "author_url": "", "post_date": "04/28/2017 23:32:18", "content": "Little hint: every sea lion female gives burth to one baby. Any sea lion pups must be always near its mother (female).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "178763": "Hey, how have you guys been handling image segmentation? I'm looking to use a sliding windows + CNN approach, but I want to have a plan for the full pipeline before implementing it. When we're using sliding windows for classification, and we've got a section of overlapping windows that all were classified as e.g. baby sea lions, how can we know how many baby sea lions are in that area? \n\nI thought about dividing the size of that area by the size of a baby sea lion, but that seems imprecise. I also thought about applying an image segmentation algorithm, but it seems like those generally require image masks in the training data. \n\nAny chance someone could point me in the right direction for this part of the challenge?", "178768": "> how can we know how many baby sea lions are in that area?\n\nOne approach would be to sum the number of locations that give baby sea lion probability above some threshold, do that for several thresholds (e.g. 0.05, 0.25, 0.5), and use that as features for a regression algorithm (e.g. xgboost or lasso) that you train on your validation set.", "178827": "Little hint: every sea lion female gives burth to one baby. Any sea lion pups must be always near its mother (female).", "178970": "Nice, that sounds like it could work. Thanks for your help!" }, "source": "meta" }