Dataset: https://www.kaggle.com/datasets/edomingo/contrails-images-ash-color-frames-3-4-5-6
\nHi everyone, following the work done by Shashwat Raman in order to generate a refined and distilled dataset (Contrails Images (Ash Color)) from the most useful frames in the entire dataset, containing only the 5th frame (the target one) in float16, here you can find an extended version of this idea with the addition of frames 3, 4 and 6. So 2 times before and one time after. Each frame has the RGB channels in False Color. The mask is stacked at the end of them.
\nSo the shape of every image here is [256 x 256 x 13]:
\n[H x W x (frame3-RGB + frame4-RGB + frame5-target-RGB + frame6-RGB + mask)]
3 RGB channels x 4 frames + 1 mask = 13 channels in the last dimension
\nI've chosen the 2 previous ones and the 1 following the target as those are apparently helping the most in improving the model's accuracy as shown in the original paper: OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI. Let me know if this is helpful and thank you @shashwatraman for your work!
", "messages": [ { "id": 2356047, "postDate": "2023-07-23T20:29:47.837Z", "content": "Dataset: https://www.kaggle.com/datasets/edomingo/contrails-images-ash-color-frames-3-4-5-6
\nHi everyone, following the work done by Shashwat Raman in order to generate a refined and distilled dataset (Contrails Images (Ash Color)) from the most useful frames in the entire dataset, containing only the 5th frame (the target one) in float16, here you can find an extended version of this idea with the addition of frames 3, 4 and 6. So 2 times before and one time after. Each frame has the RGB channels in False Color. The mask is stacked at the end of them.
\nSo the shape of every image here is [256 x 256 x 13]:
\n[H x W x (frame3-RGB + frame4-RGB + frame5-target-RGB + frame6-RGB + mask)]
3 RGB channels x 4 frames + 1 mask = 13 channels in the last dimension
\nI've chosen the 2 previous ones and the 1 following the target as those are apparently helping the most in improving the model's accuracy as shown in the original paper: OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI. Let me know if this is helpful and thank you @shashwatraman for your work!
", "rawMarkdown": "Dataset: https://www.kaggle.com/datasets/edomingo/contrails-images-ash-color-frames-3-4-5-6\n\nHi everyone, following the work done by [Shashwat Raman](https://www.kaggle.com/shashwatraman) in order to generate a refined and distilled dataset ([Contrails Images (Ash Color)](https://www.kaggle.com/datasets/shashwatraman/contrails-images-ash-color)) from the most useful frames in the entire dataset, containing only the 5th frame (the target one) in float16, here you can find an extended version of this idea with the addition of frames 3, 4 and 6. So 2 times before and one time after. Each frame has the RGB channels in False Color. The mask is stacked at the end of them.\n\nSo the shape of every image here is [256 x 256 x 13]: \n[H x W x (frame3-RGB + frame4-RGB + frame5-target-RGB + frame6-RGB + mask)]\n\n3 RGB channels x 4 frames + 1 mask = 13 channels in the last dimension\n\nI've chosen the 2 previous ones and the 1 following the target as those are apparently helping the most in improving the model's accuracy as shown in the original paper: [OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI](https://arxiv.org/abs/2304.02122). Let me know if this is helpful and thank you @shashwatraman for your work!", "votes": 11 }, { "id": 2367612, "postDate": "2023-07-31T16:59:03.297Z", "content": "Great work! May I know if there's any significant improvement in the LB score after using this dataset?
", "rawMarkdown": "Great work! May I know if there's any significant improvement in the LB score after using this dataset?", "votes": 1, "replies": [ { "id": 2367741, "postDate": "2023-07-31T19:04:40.333Z", "content": "Thanks Swaroop! We haven't tried much around it yet but apparently, from the paper's conclusions, adding the previous 2 layers could slightly improve results and also the one next to the target but with less impact than the previous two.
", "rawMarkdown": "Thanks Swaroop! We haven't tried much around it yet but apparently, from the paper's conclusions, adding the previous 2 layers could slightly improve results and also the one next to the target but with less impact than the previous two.", "votes": 1, "replies": [ { "id": 2367827, "postDate": "2023-07-31T21:06:33.117Z", "content": "I see. Thanks for your reply.
", "rawMarkdown": "I see. Thanks for your reply." } ] }, { "id": 2370262, "postDate": "2023-08-02T09:23:26.340Z", "content": "I tried a few things (Unet with Efficientnet or Resnest backbone) with the current image and 3 before as inputs as in the paper but it did not improve my results. I only used 2D models though!
", "rawMarkdown": "I tried a few things (Unet with Efficientnet or Resnest backbone) with the current image and 3 before as inputs as in the paper but it did not improve my results. I only used 2D models though!", "votes": 2, "replies": [ { "id": 2370706, "postDate": "2023-08-02T15:30:43.260Z", "content": "Got it. Thanks for sharing your findings.
", "rawMarkdown": "Got it. Thanks for sharing your findings." } ] } ] } ], "comments": [ { "id": 2367612, "author_name": "Swaroop Meher", "author_url": "", "post_date": "2023-07-31T16:59:03.297000", "content": "Great work! May I know if there's any significant improvement in the LB score after using this dataset?
", "votes": 1, "replies": [ { "id": 2367741, "author_name": "Enric Domingo", "author_url": "", "post_date": "2023-07-31T19:04:40.333000", "content": "Thanks Swaroop! We haven't tried much around it yet but apparently, from the paper's conclusions, adding the previous 2 layers could slightly improve results and also the one next to the target but with less impact than the previous two.
", "votes": 1, "replies": [ { "id": 2367827, "author_name": "Swaroop Meher", "author_url": "", "post_date": "2023-07-31T21:06:33.117000", "content": "I see. Thanks for your reply.
", "votes": 0, "replies": [] } ] }, { "id": 2370262, "author_name": "Francois Lemarchand", "author_url": "", "post_date": "2023-08-02T09:23:26.340000", "content": "I tried a few things (Unet with Efficientnet or Resnest backbone) with the current image and 3 before as inputs as in the paper but it did not improve my results. I only used 2D models though!
", "votes": 2, "replies": [ { "id": 2370706, "author_name": "Swaroop Meher", "author_url": "", "post_date": "2023-08-02T15:30:43.260000", "content": "Got it. Thanks for sharing your findings.
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2356047": "Dataset: https://www.kaggle.com/datasets/edomingo/contrails-images-ash-color-frames-3-4-5-6\n\nHi everyone, following the work done by [Shashwat Raman](https://www.kaggle.com/shashwatraman) in order to generate a refined and distilled dataset ([Contrails Images (Ash Color)](https://www.kaggle.com/datasets/shashwatraman/contrails-images-ash-color)) from the most useful frames in the entire dataset, containing only the 5th frame (the target one) in float16, here you can find an extended version of this idea with the addition of frames 3, 4 and 6. So 2 times before and one time after. Each frame has the RGB channels in False Color. The mask is stacked at the end of them.\n\nSo the shape of every image here is [256 x 256 x 13]: \n[H x W x (frame3-RGB + frame4-RGB + frame5-target-RGB + frame6-RGB + mask)]\n\n3 RGB channels x 4 frames + 1 mask = 13 channels in the last dimension\n\nI've chosen the 2 previous ones and the 1 following the target as those are apparently helping the most in improving the model's accuracy as shown in the original paper: [OpenContrails: Benchmarking Contrail Detection on GOES-16 ABI](https://arxiv.org/abs/2304.02122). Let me know if this is helpful and thank you @shashwatraman for your work!", "2367612": "Great work! May I know if there's any significant improvement in the LB score after using this dataset?" } }