The main problem that I am facing in this event is the large size of the labelled images. I am down-sampling the input images to a resolution (320,480) preserving the aspect ratio. I am using a Fully-\n Convolutional Network with Learnable Upsampling. The output is (320,480) which is then upsampled using deconvolutional layers to the (1280,1920). Given the label size is very odd (1280,1918) in he last layer I have resize the output before calculating the training loss. Is there any way i can get an accuracy near to 0.997 using the input image size i have use. Pooling the input with pseudorandom sequences added to the accuracy but the maximum i am able to reach is 99.
", "messages": [ { "id": "217695", "postDate": "08/31/2017 16:30:30", "content": "The main problem that I am facing in this event is the large size of the labelled images. I am down-sampling the input images to a resolution (320,480) preserving the aspect ratio. I am using a Fully-\n Convolutional Network with Learnable Upsampling. The output is (320,480) which is then upsampled using deconvolutional layers to the (1280,1920). Given the label size is very odd (1280,1918) in he last layer I have resize the output before calculating the training loss. Is there any way i can get an accuracy near to 0.997 using the input image size i have use. Pooling the input with pseudorandom sequences added to the accuracy but the maximum i am able to reach is 99.
", "rawMarkdown": "The main problem that I am facing in this event is the large size of the labelled images. I am down-sampling the input images to a resolution (320,480) preserving the aspect ratio. I am using a Fully-\n Convolutional Network with Learnable Upsampling. The output is (320,480) which is then upsampled using deconvolutional layers to the (1280,1920). Given the label size is very odd (1280,1918) in he last layer I have resize the output before calculating the training loss. Is there any way i can get an accuracy near to 0.997 using the input image size i have use. Pooling the input with pseudorandom sequences added to the accuracy but the maximum i am able to reach is 99.", "votes": null }, { "id": "217831", "postDate": "09/01/2017 06:30:00", "content": "With a 256x256 downscaled image (input and output) my best local scores are in the 0.996 region. I don't think 0.997 is possible with 320x480 as input and output resolution. But with your additional upscaling layers 0.997 might be doable. I did some test with replacing the final resize step from (256x256) -> 1280x1918 with some additional upsampling deconv. But it didn't work for me, results were the same or worse then simple linear resize.
", "rawMarkdown": "With a 256x256 downscaled image (input and output) my best local scores are in the 0.996 region. I don't think 0.997 is possible with 320x480 as input and output resolution. But with your additional upscaling layers 0.997 might be doable. I did some test with replacing the final resize step from (256x256) -> 1280x1918 with some additional upsampling deconv. But it didn't work for me, results were the same or worse then simple linear resize.", "votes": null }, { "id": "217846", "postDate": "09/01/2017 07:58:28", "content": "Wow, it's a pretty high score for 256, it's almost like the perfect dice score (regarding to this kernel https://www.kaggle.com/uiiurz1/how-does-the-image-scale-affect-dice the max dice score of 256 is around ~0.9965 ). Out of curiosity - does your LB is 1024?
", "rawMarkdown": "Wow, it's a pretty high score for 256, it's almost like the perfect dice score (regarding to this kernel https://www.kaggle.com/uiiurz1/how-does-the-image-scale-affect-dice the max dice score of 256 is around ~0.9965 ). Out of curiosity - does your LB is 1024?", "votes": null }, { "id": "217888", "postDate": "09/01/2017 11:31:34", "content": "Yes, my current public LB is 1024x1024. Because of rounding we are Speaking of 0.996 as 0.9955x...0.9964x I think.
", "rawMarkdown": "Yes, my current public LB is 1024x1024. Because of rounding we are Speaking of 0.996 as 0.9955x...0.9964x I think.", "votes": null } ], "comments": [ { "id": 217831, "author_name": "friedeks", "author_url": "", "post_date": "09/01/2017 06:30:00", "content": "With a 256x256 downscaled image (input and output) my best local scores are in the 0.996 region. I don't think 0.997 is possible with 320x480 as input and output resolution. But with your additional upscaling layers 0.997 might be doable. I did some test with replacing the final resize step from (256x256) -> 1280x1918 with some additional upsampling deconv. But it didn't work for me, results were the same or worse then simple linear resize.
", "votes": null, "replies": [ { "id": 217846, "author_name": "heyt0ny", "author_url": "", "post_date": "09/01/2017 07:58:28", "content": "Wow, it's a pretty high score for 256, it's almost like the perfect dice score (regarding to this kernel https://www.kaggle.com/uiiurz1/how-does-the-image-scale-affect-dice the max dice score of 256 is around ~0.9965 ). Out of curiosity - does your LB is 1024?
", "votes": null, "replies": [] }, { "id": 217888, "author_name": "friedeks", "author_url": "", "post_date": "09/01/2017 11:31:34", "content": "Yes, my current public LB is 1024x1024. Because of rounding we are Speaking of 0.996 as 0.9955x...0.9964x I think.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "217695": "The main problem that I am facing in this event is the large size of the labelled images. I am down-sampling the input images to a resolution (320,480) preserving the aspect ratio. I am using a Fully-\n Convolutional Network with Learnable Upsampling. The output is (320,480) which is then upsampled using deconvolutional layers to the (1280,1920). Given the label size is very odd (1280,1918) in he last layer I have resize the output before calculating the training loss. Is there any way i can get an accuracy near to 0.997 using the input image size i have use. Pooling the input with pseudorandom sequences added to the accuracy but the maximum i am able to reach is 99.", "217831": "With a 256x256 downscaled image (input and output) my best local scores are in the 0.996 region. I don't think 0.997 is possible with 320x480 as input and output resolution. But with your additional upscaling layers 0.997 might be doable. I did some test with replacing the final resize step from (256x256) -> 1280x1918 with some additional upsampling deconv. But it didn't work for me, results were the same or worse then simple linear resize.", "217846": "Wow, it's a pretty high score for 256, it's almost like the perfect dice score (regarding to this kernel https://www.kaggle.com/uiiurz1/how-does-the-image-scale-affect-dice the max dice score of 256 is around ~0.9965 ). Out of curiosity - does your LB is 1024?", "217888": "Yes, my current public LB is 1024x1024. Because of rounding we are Speaking of 0.996 as 0.9955x...0.9964x I think." }, "source": "meta" }