\n
![\"\"](\"https://i.ibb.co/4s4K2R1/image.png\")
there are many black and white images how are you guys gonna approach them
\n
there are many black and white images how are you guys gonna approach them
\n
just using open CV to get the images fixes them I think
", "rawMarkdown": "just using open CV to get the images fixes them I think", "votes": null }, { "id": "1685304", "postDate": "02/11/2022 07:16:25", "content": "still, that's bias possibility for the model
", "rawMarkdown": "still, that's bias possibility for the model", "votes": null }, { "id": "1685826", "postDate": "02/11/2022 15:08:22", "content": "Good insight, @somesh88, but it seems that the left bottom image is still in RGB channels.
", "rawMarkdown": "Good insight, @somesh88, but it seems that the left bottom image is still in RGB channels.", "votes": null }, { "id": "1686037", "postDate": "02/11/2022 17:54:33", "content": "What other approaches might be possible to handle this well?
\nI would just copy over the channels as @mrinath suggested.
\nOkay: A LONG Stretch-but maybe DeOldify to color the images might be a fun experiment.
", "rawMarkdown": "What other approaches might be possible to handle this well? \n\nI would just copy over the channels as @mrinath suggested. \n\nOkay: A LONG Stretch-but maybe DeOldify to color the images might be a fun experiment.", "votes": null }, { "id": "1688452", "postDate": "02/13/2022 16:19:09", "content": "Thanks for the visualizations @somesh88! I have analyzed the distribution of single-channel images for all test and train images in my notebook here link to the relevant section. You may find it interesting!
", "rawMarkdown": "Thanks for the visualizations @somesh88! I have analyzed the distribution of single-channel images for all test and train images in my notebook here [link to the relevant section](https://www.kaggle.com/bsridatta/happywhale/notebook#Sample-distributions-%F0%9F%93%88). You may find it interesting!", "votes": null }, { "id": "1688479", "postDate": "02/13/2022 16:46:17", "content": "yes it was 4 channeled image I just checked if the image is not 3 channeled haha
", "rawMarkdown": "yes it was 4 channeled image I just checked if the image is not 3 channeled haha", "votes": null }, { "id": "1688485", "postDate": "02/13/2022 16:50:05", "content": "hey great notebook thanks for sharing it with me :)
", "rawMarkdown": "hey great notebook thanks for sharing it with me :)", "votes": null }, { "id": "1688487", "postDate": "02/13/2022 16:51:29", "content": "haha yes it'll be a fun experiment but don't have that many compute resources available rn ( Kaggle provides less GPU hrs :( )
", "rawMarkdown": "haha yes it'll be a fun experiment but don't have that many compute resources available rn ( Kaggle provides less GPU hrs :( )", "votes": null }, { "id": "1688488", "postDate": "02/13/2022 16:52:14", "content": "I've seen one notebook he was using cv2 it was same copying over the channels approach haha
", "rawMarkdown": "I've seen one notebook he was using cv2 it was same copying over the channels approach haha", "votes": null }, { "id": "1688510", "postDate": "02/13/2022 17:04:56", "content": "I think DeOldify uses really fewer resources. It might be worth trying on Colab.
\nIt would also make for an interesting dataset if anyone is willing to experiment :)
", "rawMarkdown": "I think DeOldify uses really fewer resources. It might be worth trying on Colab. \n\nIt would also make for an interesting dataset if anyone is willing to experiment :)", "votes": null }, { "id": "1688517", "postDate": "02/13/2022 17:11:03", "content": "I'll try it on Kaggle let's see how it goes
\ncreating notebook right now doesn't seems that hard ( but I am sure finding the 3 channeled images takes lot of time lol )
@init27 the iamge data is of around 62 gb not sure I can make it in kaggle directory :(
", "rawMarkdown": "init27 the iamge data is of around 62 gb not sure I can make it in kaggle directory :(", "votes": null }, { "id": "1688566", "postDate": "02/13/2022 17:46:08", "content": "@init27 I was trying out deoldify but dataset is too big I am not sure how long it will take but I made some visualizations what do you think will it be helpful to convert these 1channeld images to colorized images
\n![\"\"](\"https://i.ibb.co/kKT3H3N/image.png\")
some more images here ![\"\"](\"https://i.ibb.co/bQGF5DG/image.png\")
It's a good idea to add this to the title or main content, I am sure many didn't notice this. I suspected something like this and handled it in creating meta data but was too lazy to verify how many are like that. Thanks for your insight!
", "rawMarkdown": "It's a good idea to add this to the title or main content, I am sure many didn't notice this. I suspected something like this and handled it in creating meta data but was too lazy to verify how many are like that. Thanks for your insight!", "votes": null }, { "id": "1688662", "postDate": "02/13/2022 18:40:42", "content": "I dont think you should worry about that. They don't visually look different from the 3 channel ones after converting to RGB in many cases. And the proportion is very small.
", "rawMarkdown": "I dont think you should worry about that. They don't visually look different from the 3 channel ones after converting to RGB in many cases. And the proportion is very small.", "votes": null }, { "id": "1688665", "postDate": "02/13/2022 18:45:36", "content": "no it's not small in first 1000 images I found about 24 images b_and_w
\nin total 54000 I think there will be hell lot of images
Please look at how competitors batch data-there have been many similar competitions.
\nYou'll have to work with subsets.
", "rawMarkdown": "Please look at how competitors batch data-there have been many similar competitions. \n\nYou'll have to work with subsets.", "votes": null }, { "id": "1689659", "postDate": "02/14/2022 11:48:07", "content": "@init27 the dataset is published now thanks for your idea
\nthere are images but I manged to generate colorized images haha
https://www.kaggle.com/somesh88/coloring-single-channel-images-deoldify?scriptVersionId=87787281
\nhttps://www.kaggle.com/somesh88/deoldify-single-channeled-converted-images
", "rawMarkdown": "init27 the dataset is published now thanks for your idea \nthere are images but I manged to generate colorized images haha \n\n## completed notebook link :\n[https://www.kaggle.com/somesh88/coloring-single-channel-images-deoldify?scriptVersionId=87787281](https://www.kaggle.com/somesh88/coloring-single-channel-images-deoldify?scriptVersionId=87787281)\n## dataset link \n[https://www.kaggle.com/somesh88/deoldify-single-channeled-converted-images](https://www.kaggle.com/somesh88/deoldify-single-channeled-converted-images)", "votes": null } ], "comments": [ { "id": 1685267, "author_name": "mrinath", "author_url": "", "post_date": "02/11/2022 06:40:46", "content": "just using open CV to get the images fixes them I think
", "votes": null, "replies": [ { "id": 1685304, "author_name": "somesh88", "author_url": "", "post_date": "02/11/2022 07:16:25", "content": "still, that's bias possibility for the model
", "votes": null, "replies": [] }, { "id": 1686037, "author_name": "init27", "author_url": "", "post_date": "02/11/2022 17:54:33", "content": "What other approaches might be possible to handle this well?
\nI would just copy over the channels as @mrinath suggested.
\nOkay: A LONG Stretch-but maybe DeOldify to color the images might be a fun experiment.
", "votes": null, "replies": [] }, { "id": 1688487, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 16:51:29", "content": "haha yes it'll be a fun experiment but don't have that many compute resources available rn ( Kaggle provides less GPU hrs :( )
", "votes": null, "replies": [] }, { "id": 1688488, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 16:52:14", "content": "I've seen one notebook he was using cv2 it was same copying over the channels approach haha
", "votes": null, "replies": [] }, { "id": 1688510, "author_name": "init27", "author_url": "", "post_date": "02/13/2022 17:04:56", "content": "I think DeOldify uses really fewer resources. It might be worth trying on Colab.
\nIt would also make for an interesting dataset if anyone is willing to experiment :)
", "votes": null, "replies": [] }, { "id": 1688517, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 17:11:03", "content": "I'll try it on Kaggle let's see how it goes
\ncreating notebook right now doesn't seems that hard ( but I am sure finding the 3 channeled images takes lot of time lol )
@init27 the iamge data is of around 62 gb not sure I can make it in kaggle directory :(
", "votes": null, "replies": [] }, { "id": 1688745, "author_name": "init27", "author_url": "", "post_date": "02/13/2022 19:46:57", "content": "Please look at how competitors batch data-there have been many similar competitions.
\nYou'll have to work with subsets.
", "votes": null, "replies": [] }, { "id": 1689659, "author_name": "somesh88", "author_url": "", "post_date": "02/14/2022 11:48:07", "content": "@init27 the dataset is published now thanks for your idea
\nthere are images but I manged to generate colorized images haha
https://www.kaggle.com/somesh88/coloring-single-channel-images-deoldify?scriptVersionId=87787281
\nhttps://www.kaggle.com/somesh88/deoldify-single-channeled-converted-images
", "votes": null, "replies": [] } ] }, { "id": 1685826, "author_name": "vad13irt", "author_url": "", "post_date": "02/11/2022 15:08:22", "content": "Good insight, @somesh88, but it seems that the left bottom image is still in RGB channels.
", "votes": null, "replies": [ { "id": 1688479, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 16:46:17", "content": "yes it was 4 channeled image I just checked if the image is not 3 channeled haha
", "votes": null, "replies": [] }, { "id": 1688655, "author_name": "bsridatta", "author_url": "", "post_date": "02/13/2022 18:38:49", "content": "It's a good idea to add this to the title or main content, I am sure many didn't notice this. I suspected something like this and handled it in creating meta data but was too lazy to verify how many are like that. Thanks for your insight!
", "votes": null, "replies": [] } ] }, { "id": 1688452, "author_name": "bsridatta", "author_url": "", "post_date": "02/13/2022 16:19:09", "content": "Thanks for the visualizations @somesh88! I have analyzed the distribution of single-channel images for all test and train images in my notebook here link to the relevant section. You may find it interesting!
", "votes": null, "replies": [ { "id": 1688485, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 16:50:05", "content": "hey great notebook thanks for sharing it with me :)
", "votes": null, "replies": [] } ] }, { "id": 1688566, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 17:46:08", "content": "@init27 I was trying out deoldify but dataset is too big I am not sure how long it will take but I made some visualizations what do you think will it be helpful to convert these 1channeld images to colorized images
\n
some more images here
I dont think you should worry about that. They don't visually look different from the 3 channel ones after converting to RGB in many cases. And the proportion is very small.
", "votes": null, "replies": [] }, { "id": 1688665, "author_name": "somesh88", "author_url": "", "post_date": "02/13/2022 18:45:36", "content": "no it's not small in first 1000 images I found about 24 images b_and_w
\nin total 54000 I think there will be hell lot of images