Hi, I am a new one in object detection. And I am trying to use centernet as my algorithm. I follow this kernel centernet-baseline , and use the same loss function, i.e. entropy function of mask and L1 loss of 6d regression. The mask tells us where the cars are in the image. The difference is that he uses efficientnet and I use hourglass.
\n\nNow my network works very well predicting the mask, indicating it learns well the 2d coordinates of cars in the image. But the 6d regression is very bad. I tried to increase the coefficient of L1 loss, and it works a little bit, but the result is still bad.
\n\nAny ideas to improve accuracy on regression? Like transforming 6d coordinates to some other things easier for the network to predict?
", "messages": [ { "id": 689606, "postDate": "2019-12-07T06:38:07.880Z", "content": "Hi, I am a new one in object detection. And I am trying to use centernet as my algorithm. I follow this kernel centernet-baseline , and use the same loss function, i.e. entropy function of mask and L1 loss of 6d regression. The mask tells us where the cars are in the image. The difference is that he uses efficientnet and I use hourglass.
\n\nNow my network works very well predicting the mask, indicating it learns well the 2d coordinates of cars in the image. But the 6d regression is very bad. I tried to increase the coefficient of L1 loss, and it works a little bit, but the result is still bad.
\n\nAny ideas to improve accuracy on regression? Like transforming 6d coordinates to some other things easier for the network to predict?
", "rawMarkdown": "Hi, I am a new one in object detection. And I am trying to use centernet as my algorithm. I follow this kernel [centernet-baseline](https://www.kaggle.com/hocop1/centernet-baseline) , and use the same loss function, i.e. entropy function of mask and L1 loss of 6d regression. The mask tells us where the cars are in the image. The difference is that he uses efficientnet and I use hourglass. \n\nNow my network works very well predicting the mask, indicating it learns well the 2d coordinates of cars in the image. But the 6d regression is very bad. I tried to increase the coefficient of L1 loss, and it works a little bit, but the result is still bad.\n\nAny ideas to improve accuracy on regression? Like transforming 6d coordinates to some other things easier for the network to predict?", "votes": 2 }, { "id": 689928, "postDate": "2019-12-07T17:29:57.440Z", "rawMarkdown": "", "replies": [ { "id": 701331, "postDate": "2019-12-23T11:18:47.213Z", "content": "May i ask what's the input image size(model input size) of train and test? @diegojohnson
", "rawMarkdown": "May i ask what's the input image size(model input size) of train and test? @diegojohnson " }, { "id": 701358, "postDate": "2019-12-23T11:53:28.090Z", "content": "It's up to you. @cswwp347724 \nyou can refer to public notebook.
", "rawMarkdown": "It's up to you. @cswwp347724 \nyou can refer to public notebook." }, { "id": 701910, "postDate": "2019-12-24T03:25:48.677Z", "content": "@diegojohnson I've tried 512*512, result seems not good, lb just 0.03
", "rawMarkdown": "@diegojohnson I've tried 512*512, result seems not good, lb just 0.03" } ] }, { "id": 689937, "postDate": "2019-12-07T17:41:38.497Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 689928, "author_name": "DiegoJohnson", "author_url": "", "post_date": "2019-12-07T17:29:57.440000", "content": "", "votes": 0, "replies": [ { "id": 701331, "author_name": "cswwp", "author_url": "", "post_date": "2019-12-23T11:18:47.213000", "content": "May i ask what's the input image size(model input size) of train and test? @diegojohnson
", "votes": 0, "replies": [] }, { "id": 701358, "author_name": "DiegoJohnson", "author_url": "", "post_date": "2019-12-23T11:53:28.090000", "content": "It's up to you. @cswwp347724 \nyou can refer to public notebook.
", "votes": 0, "replies": [] }, { "id": 701910, "author_name": "cswwp", "author_url": "", "post_date": "2019-12-24T03:25:48.677000", "content": "@diegojohnson I've tried 512*512, result seems not good, lb just 0.03
", "votes": 0, "replies": [] } ] }, { "id": 689937, "author_name": "", "author_url": "", "post_date": "2019-12-07T17:41:38.497000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "689606": "Hi, I am a new one in object detection. And I am trying to use centernet as my algorithm. I follow this kernel [centernet-baseline](https://www.kaggle.com/hocop1/centernet-baseline) , and use the same loss function, i.e. entropy function of mask and L1 loss of 6d regression. The mask tells us where the cars are in the image. The difference is that he uses efficientnet and I use hourglass. \n\nNow my network works very well predicting the mask, indicating it learns well the 2d coordinates of cars in the image. But the 6d regression is very bad. I tried to increase the coefficient of L1 loss, and it works a little bit, but the result is still bad.\n\nAny ideas to improve accuracy on regression? Like transforming 6d coordinates to some other things easier for the network to predict?", "689928": "", "689937": "" } }