I am the new participant of this competition and I am the newbie of deep learning....\nI know it's close to deadline, just curious why there are no nay kernel use MRCNN?\nI seem there are many kernels use Unet to do this competitions. \nIt's because that use the kernel to train M-RCNN will dead or there is something else reason.
", "messages": [ { "id": 414244, "postDate": "2018-11-02T11:51:10.073Z", "content": "I think mots of people didn't choose mask r-cnn.
\n\nFirst, The dataset of mask r-cnn is difficult to generate. I spent quite long time to generate dataset. You can see here.
\n\nSecond, instance segmentation is different from semantic segmentation, wich will lead to some problems in exporting results. If you go on with Mask R-CNN, you will find these difficulities. They are difficult to deal with.
", "rawMarkdown": "I think mots of people didn't choose mask r-cnn. \n\nFirst, The dataset of mask r-cnn is difficult to generate. I spent quite long time to generate dataset. You can see [here](https://www.kaggle.com/c/airbus-ship-detection/discussion/68790#latest-414204).\n\nSecond, instance segmentation is different from semantic segmentation, wich will lead to some problems in exporting results. If you go on with Mask R-CNN, you will find these difficulities. They are difficult to deal with.", "votes": 1, "replies": [ { "id": 414321, "postDate": "2018-11-02T14:41:20.920Z", "content": "Thanks!! No wonder there are no one use Mask R-CNN in this competition.
", "rawMarkdown": "Thanks!! No wonder there are no one use Mask R-CNN in this competition." }, { "id": 414337, "postDate": "2018-11-02T15:15:58.377Z", "content": "In fact, there are a few people using Mask R-CNN, i konw several people use detectron, but they are not willing to share kernels. In addition, the environment of detectron is complex, i'm not sure whether our codes can be run on kernel.
", "rawMarkdown": "In fact, there are a few people using Mask R-CNN, i konw several people use detectron, but they are not willing to share kernels. In addition, the environment of detectron is complex, i'm not sure whether our codes can be run on kernel." } ] }, { "id": 413818, "postDate": "2018-11-01T15:45:53.917Z", "content": "I am the new participant of this competition and I am the newbie of deep learning....\nI know it's close to deadline, just curious why there are no nay kernel use MRCNN?\nI seem there are many kernels use Unet to do this competitions. \nIt's because that use the kernel to train M-RCNN will dead or there is something else reason.
", "rawMarkdown": "I am the new participant of this competition and I am the newbie of deep learning....\nI know it's close to deadline, just curious why there are no nay kernel use MRCNN?\nI seem there are many kernels use Unet to do this competitions. \nIt's because that use the kernel to train M-RCNN will dead or there is something else reason.", "votes": 2 }, { "id": 414974, "postDate": "2018-11-04T02:08:19.047Z", "content": "I tried to train a M-RCNN for this competition with very terrible results. First, it took much longer to converge than unet. I trained the NN for 50 epochs for two days and the score was worse than the empty submission. I also saw a lot of false positives, and then I stacked RCNN with a model that detects ship/no ship and it did slightly better than the empty submission but not by much.
", "rawMarkdown": "I tried to train a M-RCNN for this competition with very terrible results. First, it took much longer to converge than unet. I trained the NN for 50 epochs for two days and the score was worse than the empty submission. I also saw a lot of false positives, and then I stacked RCNN with a model that detects ship/no ship and it did slightly better than the empty submission but not by much.", "replies": [ { "id": 415017, "postDate": "2018-11-04T06:20:08.077Z", "content": "Can you share the score of empty submission? I remember that someone said it means the ratio of empty images.
", "rawMarkdown": " Can you share the score of empty submission? I remember that [someone](https://www.kaggle.com/c/airbus-ship-detection/discussion/62376) said it means the ratio of empty images." }, { "id": 415544, "postDate": "2018-11-05T09:49:23.430Z", "content": "Hi Pascal \nYou can refer this kernel : https://www.kaggle.com/jeffaudi/create-a-submission-with-no-ship-masks\nI think the empty submit score : 0.52
", "rawMarkdown": "Hi Pascal \nYou can refer this kernel : https://www.kaggle.com/jeffaudi/create-a-submission-with-no-ship-masks\nI think the empty submit score : 0.52" }, { "id": 415545, "postDate": "2018-11-05T09:50:35.293Z", "content": "Thanks，it seems not good.
", "rawMarkdown": "Thanks，it seems not good." } ] }, { "id": 414647, "postDate": "2018-11-03T09:30:41.757Z", "content": "We too use MaskRCNN, Well there are kernels available in Kaggle but not in the competition's I guess. Like you can see one here: https://www.kaggle.com/drt2290078/mask-rcnn-sample-starter-code .
", "rawMarkdown": "We too use MaskRCNN, Well there are kernels available in Kaggle but not in the competition's I guess. Like you can see one here: https://www.kaggle.com/drt2290078/mask-rcnn-sample-starter-code . " }, { "id": 414575, "postDate": "2018-11-03T04:12:31.440Z", "content": "My current solution is based on MaskRCNN. But the pipeline is more complex than Unet based solutions and harder to improve. I am thinking to move to Unet based solutions :( .
", "rawMarkdown": "My current solution is based on MaskRCNN. But the pipeline is more complex than Unet based solutions and harder to improve. I am thinking to move to Unet based solutions :( .", "replies": [ { "id": 414597, "postDate": "2018-11-03T05:56:04.600Z", "rawMarkdown": "", "isDeleted": true }, { "id": 421629, "postDate": "2018-11-15T08:03:27.570Z", "content": "what score did you get from Mask-RCNN? we got an aweful 0.38 results which is much much worse from unet baseline (~0.6)
", "rawMarkdown": "what score did you get from Mask-RCNN? we got an aweful 0.38 results which is much much worse from unet baseline (~0.6)" }, { "id": 421632, "postDate": "2018-11-15T08:10:28.790Z", "content": "Detectron + Unet
", "rawMarkdown": "Detectron + Unet" } ] } ], "comments": [ { "id": 414244, "author_name": "pascal1129", "author_url": "", "post_date": "2018-11-02T11:51:10.073000", "content": "I think mots of people didn't choose mask r-cnn.
\n\nFirst, The dataset of mask r-cnn is difficult to generate. I spent quite long time to generate dataset. You can see here.
\n\nSecond, instance segmentation is different from semantic segmentation, wich will lead to some problems in exporting results. If you go on with Mask R-CNN, you will find these difficulities. They are difficult to deal with.
", "votes": 1, "replies": [ { "id": 414321, "author_name": "0xFunky", "author_url": "", "post_date": "2018-11-02T14:41:20.920000", "content": "Thanks!! No wonder there are no one use Mask R-CNN in this competition.
", "votes": 0, "replies": [] }, { "id": 414337, "author_name": "pascal1129", "author_url": "", "post_date": "2018-11-02T15:15:58.377000", "content": "In fact, there are a few people using Mask R-CNN, i konw several people use detectron, but they are not willing to share kernels. In addition, the environment of detectron is complex, i'm not sure whether our codes can be run on kernel.
", "votes": 0, "replies": [] } ] }, { "id": 414974, "author_name": "Maria Wellen", "author_url": "", "post_date": "2018-11-04T02:08:19.047000", "content": "I tried to train a M-RCNN for this competition with very terrible results. First, it took much longer to converge than unet. I trained the NN for 50 epochs for two days and the score was worse than the empty submission. I also saw a lot of false positives, and then I stacked RCNN with a model that detects ship/no ship and it did slightly better than the empty submission but not by much.
", "votes": 0, "replies": [ { "id": 415017, "author_name": "pascal1129", "author_url": "", "post_date": "2018-11-04T06:20:08.077000", "content": "Can you share the score of empty submission? I remember that someone said it means the ratio of empty images.
", "votes": 0, "replies": [] }, { "id": 415544, "author_name": "0xFunky", "author_url": "", "post_date": "2018-11-05T09:49:23.430000", "content": "Hi Pascal \nYou can refer this kernel : https://www.kaggle.com/jeffaudi/create-a-submission-with-no-ship-masks\nI think the empty submit score : 0.52
", "votes": 0, "replies": [] }, { "id": 415545, "author_name": "pascal1129", "author_url": "", "post_date": "2018-11-05T09:50:35.293000", "content": "Thanks，it seems not good.
", "votes": 0, "replies": [] } ] }, { "id": 414647, "author_name": "Vaghawan ", "author_url": "", "post_date": "2018-11-03T09:30:41.757000", "content": "We too use MaskRCNN, Well there are kernels available in Kaggle but not in the competition's I guess. Like you can see one here: https://www.kaggle.com/drt2290078/mask-rcnn-sample-starter-code .
", "votes": 0, "replies": [] }, { "id": 414575, "author_name": "tkuanlun350", "author_url": "", "post_date": "2018-11-03T04:12:31.440000", "content": "My current solution is based on MaskRCNN. But the pipeline is more complex than Unet based solutions and harder to improve. I am thinking to move to Unet based solutions :( .
", "votes": 0, "replies": [ { "id": 414597, "author_name": "", "author_url": "", "post_date": "2018-11-03T05:56:04.600000", "content": "", "votes": 0, "replies": [] }, { "id": 421629, "author_name": "WongSherman", "author_url": "", "post_date": "2018-11-15T08:03:27.570000", "content": "what score did you get from Mask-RCNN? we got an aweful 0.38 results which is much much worse from unet baseline (~0.6)
", "votes": 0, "replies": [] }, { "id": 421632, "author_name": "pascal1129", "author_url": "", "post_date": "2018-11-15T08:10:28.790000", "content": "Detectron + Unet
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "414244": "I think mots of people didn't choose mask r-cnn. \n\nFirst, The dataset of mask r-cnn is difficult to generate. I spent quite long time to generate dataset. You can see [here](https://www.kaggle.com/c/airbus-ship-detection/discussion/68790#latest-414204).\n\nSecond, instance segmentation is different from semantic segmentation, wich will lead to some problems in exporting results. If you go on with Mask R-CNN, you will find these difficulities. They are difficult to deal with.", "413818": "I am the new participant of this competition and I am the newbie of deep learning....\nI know it's close to deadline, just curious why there are no nay kernel use MRCNN?\nI seem there are many kernels use Unet to do this competitions. \nIt's because that use the kernel to train M-RCNN will dead or there is something else reason.", "414974": "I tried to train a M-RCNN for this competition with very terrible results. First, it took much longer to converge than unet. I trained the NN for 50 epochs for two days and the score was worse than the empty submission. I also saw a lot of false positives, and then I stacked RCNN with a model that detects ship/no ship and it did slightly better than the empty submission but not by much.", "414647": "We too use MaskRCNN, Well there are kernels available in Kaggle but not in the competition's I guess. Like you can see one here: https://www.kaggle.com/drt2290078/mask-rcnn-sample-starter-code . ", "414575": "My current solution is based on MaskRCNN. But the pipeline is more complex than Unet based solutions and harder to improve. I am thinking to move to Unet based solutions :( ." } }