Hello everybody:\nCan anyone explains to me, how is it possible to do an instance-level semantic segmentation, without having the ground truth bounding boxes for the objects; since almost all the algorithms like Mask-RCNN and many others, are depending on the object's predicted bounding box to resize the predicting mask and to localize it in the original image.\nThank you very much for sharing your ideas and I'm sorry if I misunderstand any part of this competition.
", "messages": [ { "id": "334287", "postDate": "05/26/2018 21:56:08", "content": "Hello everybody:\nCan anyone explains to me, how is it possible to do an instance-level semantic segmentation, without having the ground truth bounding boxes for the objects; since almost all the algorithms like Mask-RCNN and many others, are depending on the object's predicted bounding box to resize the predicting mask and to localize it in the original image.\nThank you very much for sharing your ideas and I'm sorry if I misunderstand any part of this competition.
", "rawMarkdown": "Hello everybody:\nCan anyone explains to me, how is it possible to do an instance-level semantic segmentation, without having the ground truth bounding boxes for the objects; since almost all the algorithms like Mask-RCNN and many others, are depending on the object's predicted bounding box to resize the predicting mask and to localize it in the original image.\nThank you very much for sharing your ideas and I'm sorry if I misunderstand any part of this competition.", "votes": null }, { "id": "334293", "postDate": "05/26/2018 22:11:49", "content": "You do have a mask for every instance. \n\"For example, a pixel value of 33000 means it belongs to label 33 (a car), is instance #0, while the pixel value of 33001 means it also belongs to class 33 (a car) , and is instance #1. These represent two different cars in an image.\"
", "rawMarkdown": "You do have a mask for every instance. \n\"For example, a pixel value of 33000 means it belongs to label 33 (a car), is instance #0, while the pixel value of 33001 means it also belongs to class 33 (a car) , and is instance #1. These represent two different cars in an image.\"", "votes": null }, { "id": "334296", "postDate": "05/26/2018 22:19:20", "content": "Yeah thanks for your answer, but how this will gonna be useful for localizing the objects in the frames?
", "rawMarkdown": "Yeah thanks for your answer, but how this will gonna be useful for localizing the objects in the frames?", "votes": null }, { "id": "334298", "postDate": "05/26/2018 22:22:30", "content": "If you have an instance's mask what else do you need?
", "rawMarkdown": "If you have an instance's mask what else do you need?", "votes": null }, { "id": "336667", "postDate": "06/01/2018 04:18:50", "content": "Hi Azat, if you don't mind can elaborate. Yes, I do get it, finding the instance numbers and corresponding classes. I've tried Fast-RCNN earlier and it requires coordinates of bounding boxes as @Malek pointed out. Yes, as you say since we have pixel locations of separate instances we can generate bounding boxes on our own though it would be a tedious task. Do you have suggestions to process the data?
", "rawMarkdown": "Hi Azat, if you don't mind can elaborate. Yes, I do get it, finding the instance numbers and corresponding classes. I've tried Fast-RCNN earlier and it requires coordinates of bounding boxes as @Malek pointed out. Yes, as you say since we have pixel locations of separate instances we can generate bounding boxes on our own though it would be a tedious task. Do you have suggestions to process the data?", "votes": null }, { "id": "336937", "postDate": "06/01/2018 15:13:01", "content": "@VikramanK I would like to know how to generate bounding boxes using pixel locations.
", "rawMarkdown": "VikramanK I would like to know how to generate bounding boxes using pixel locations.", "votes": null }, { "id": "337516", "postDate": "06/03/2018 04:19:20", "content": "hi @William,\nI am not sure how to do that. However, after browsing through few Kernels in DSB-18 I think we need to use Open CV helper function to find the Rectangles and then store the coordinates of this box. Nevertheless, I think this would be a tedious process in our case because of huge dataset and number of classes and instances. That's why I don't see any Kernels that apply Mask RCNN algorithm for this competition
", "rawMarkdown": "hi @William,\nI am not sure how to do that. However, after browsing through few Kernels in DSB-18 I think we need to use Open CV helper function to find the Rectangles and then store the coordinates of this box. Nevertheless, I think this would be a tedious process in our case because of huge dataset and number of classes and instances. That's why I don't see any Kernels that apply Mask RCNN algorithm for this competition", "votes": null } ], "comments": [ { "id": 334293, "author_name": "azakhtyamov", "author_url": "", "post_date": "05/26/2018 22:11:49", "content": "You do have a mask for every instance. \n\"For example, a pixel value of 33000 means it belongs to label 33 (a car), is instance #0, while the pixel value of 33001 means it also belongs to class 33 (a car) , and is instance #1. These represent two different cars in an image.\"
", "votes": null, "replies": [ { "id": 334296, "author_name": "drmalek", "author_url": "", "post_date": "05/26/2018 22:19:20", "content": "Yeah thanks for your answer, but how this will gonna be useful for localizing the objects in the frames?
", "votes": null, "replies": [] }, { "id": 334298, "author_name": "azakhtyamov", "author_url": "", "post_date": "05/26/2018 22:22:30", "content": "If you have an instance's mask what else do you need?
", "votes": null, "replies": [] }, { "id": 336667, "author_name": "vikramank", "author_url": "", "post_date": "06/01/2018 04:18:50", "content": "Hi Azat, if you don't mind can elaborate. Yes, I do get it, finding the instance numbers and corresponding classes. I've tried Fast-RCNN earlier and it requires coordinates of bounding boxes as @Malek pointed out. Yes, as you say since we have pixel locations of separate instances we can generate bounding boxes on our own though it would be a tedious task. Do you have suggestions to process the data?
", "votes": null, "replies": [] }, { "id": 336937, "author_name": "dskswu", "author_url": "", "post_date": "06/01/2018 15:13:01", "content": "@VikramanK I would like to know how to generate bounding boxes using pixel locations.
", "votes": null, "replies": [] }, { "id": 337516, "author_name": "vikramank", "author_url": "", "post_date": "06/03/2018 04:19:20", "content": "hi @William,\nI am not sure how to do that. However, after browsing through few Kernels in DSB-18 I think we need to use Open CV helper function to find the Rectangles and then store the coordinates of this box. Nevertheless, I think this would be a tedious process in our case because of huge dataset and number of classes and instances. That's why I don't see any Kernels that apply Mask RCNN algorithm for this competition
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "334287": "Hello everybody:\nCan anyone explains to me, how is it possible to do an instance-level semantic segmentation, without having the ground truth bounding boxes for the objects; since almost all the algorithms like Mask-RCNN and many others, are depending on the object's predicted bounding box to resize the predicting mask and to localize it in the original image.\nThank you very much for sharing your ideas and I'm sorry if I misunderstand any part of this competition.", "334293": "You do have a mask for every instance. \n\"For example, a pixel value of 33000 means it belongs to label 33 (a car), is instance #0, while the pixel value of 33001 means it also belongs to class 33 (a car) , and is instance #1. These represent two different cars in an image.\"", "334296": "Yeah thanks for your answer, but how this will gonna be useful for localizing the objects in the frames?", "334298": "If you have an instance's mask what else do you need?", "336667": "Hi Azat, if you don't mind can elaborate. Yes, I do get it, finding the instance numbers and corresponding classes. I've tried Fast-RCNN earlier and it requires coordinates of bounding boxes as @Malek pointed out. Yes, as you say since we have pixel locations of separate instances we can generate bounding boxes on our own though it would be a tedious task. Do you have suggestions to process the data?", "336937": "VikramanK I would like to know how to generate bounding boxes using pixel locations.", "337516": "hi @William,\nI am not sure how to do that. However, after browsing through few Kernels in DSB-18 I think we need to use Open CV helper function to find the Rectangles and then store the coordinates of this box. Nevertheless, I think this would be a tedious process in our case because of huge dataset and number of classes and instances. That's why I don't see any Kernels that apply Mask RCNN algorithm for this competition" }, "source": "meta" }