I recently came across several discussion posts centered around removing hairs from lesion images, while there were also a considerable number of posts revolving around adding hairs in an image via some pre processing technique.
\n\nWhat's the point of adding hairs?? Won't it complicate the matter?? What's the motivation behind this?
", "messages": [ { "id": "947795", "postDate": "07/27/2020 13:40:35", "content": "I recently came across several discussion posts centered around removing hairs from lesion images, while there were also a considerable number of posts revolving around adding hairs in an image via some pre processing technique.
\n\nWhat's the point of adding hairs?? Won't it complicate the matter?? What's the motivation behind this?
", "rawMarkdown": "I recently came across several discussion posts centered around removing hairs from lesion images, while there were also a considerable number of posts revolving around adding hairs in an image via some pre processing technique.\n\nWhat's the point of adding hairs?? Won't it complicate the matter?? What's the motivation behind this?", "votes": null }, { "id": "947837", "postDate": "07/27/2020 14:06:30", "content": "Aditya, some participants reported higher scores by adding hair, instead of removing it.
", "rawMarkdown": "Aditya, some participants [reported higher scores](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165582#922706) by adding hair, instead of removing it.", "votes": null }, { "id": "948342", "postDate": "07/27/2020 20:34:35", "content": "Adding hair is just a data augmentation technique applied to images on the training stage to help a model learn.
\n\nLike every other augmentation technique, the goal is to create different versions of the image such that the label is the same but the image itself is a bit different. In general, there are many techniques to do that such as cropping, cutout, rotation and others. In this competition, many images have some amount of hair on them, so adding artificial hair is a suitable augmentation for our context that is a bit similar to cutout.
\n\nYou are right that adding hair complicates the classification task. But it also motivates a model to learn patterns in the areas not covered by hair. During training, we are learning from the same image multiple times, adding differnt artificial hair in different places on each epoch. This could help to fight overfitting and eventually lead to better performance.
", "rawMarkdown": "Adding hair is just a data augmentation technique applied to images on the training stage to help a model learn.\n\nLike every other augmentation technique, the goal is to create different versions of the image such that the label is the same but the image itself is a bit different. In general, there are many techniques to do that such as cropping, cutout, rotation and others. In this competition, many images have some amount of hair on them, so adding artificial hair is a suitable augmentation for our context that is a bit similar to cutout.\n\nYou are right that adding hair complicates the classification task. But it also motivates a model to learn patterns in the areas not covered by hair. During training, we are learning from the same image multiple times, adding differnt artificial hair in different places on each epoch. This could help to fight overfitting and eventually lead to better performance.", "votes": null }, { "id": "948725", "postDate": "07/28/2020 07:03:43", "content": "Thank you for a detailed response!
", "rawMarkdown": "Thank you for a detailed response!", "votes": null }, { "id": "949621", "postDate": "07/28/2020 18:47:41", "content": "True, I can confirm. The idea seems good but removing hair lowered my LB Score significantly. Its better to stick with adding hair.
", "rawMarkdown": "True, I can confirm. The idea seems good but removing hair lowered my LB Score significantly. Its better to stick with adding hair.", "votes": null }, { "id": "950154", "postDate": "07/29/2020 08:02:32", "content": "okay, thank you!
", "rawMarkdown": "okay, thank you!", "votes": null } ], "comments": [ { "id": 947837, "author_name": "sirishks", "author_url": "", "post_date": "07/27/2020 14:06:30", "content": "Aditya, some participants reported higher scores by adding hair, instead of removing it.
", "votes": null, "replies": [ { "id": 949621, "author_name": "aliabdin1", "author_url": "", "post_date": "07/28/2020 18:47:41", "content": "True, I can confirm. The idea seems good but removing hair lowered my LB Score significantly. Its better to stick with adding hair.
", "votes": null, "replies": [] }, { "id": 950154, "author_name": "fireheart7", "author_url": "", "post_date": "07/29/2020 08:02:32", "content": "okay, thank you!
", "votes": null, "replies": [] } ] }, { "id": 948342, "author_name": "kozodoi", "author_url": "", "post_date": "07/27/2020 20:34:35", "content": "Adding hair is just a data augmentation technique applied to images on the training stage to help a model learn.
\n\nLike every other augmentation technique, the goal is to create different versions of the image such that the label is the same but the image itself is a bit different. In general, there are many techniques to do that such as cropping, cutout, rotation and others. In this competition, many images have some amount of hair on them, so adding artificial hair is a suitable augmentation for our context that is a bit similar to cutout.
\n\nYou are right that adding hair complicates the classification task. But it also motivates a model to learn patterns in the areas not covered by hair. During training, we are learning from the same image multiple times, adding differnt artificial hair in different places on each epoch. This could help to fight overfitting and eventually lead to better performance.
", "votes": null, "replies": [ { "id": 948725, "author_name": "fireheart7", "author_url": "", "post_date": "07/28/2020 07:03:43", "content": "Thank you for a detailed response!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "947795": "I recently came across several discussion posts centered around removing hairs from lesion images, while there were also a considerable number of posts revolving around adding hairs in an image via some pre processing technique.\n\nWhat's the point of adding hairs?? Won't it complicate the matter?? What's the motivation behind this?", "947837": "Aditya, some participants [reported higher scores](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/165582#922706) by adding hair, instead of removing it.", "948342": "Adding hair is just a data augmentation technique applied to images on the training stage to help a model learn.\n\nLike every other augmentation technique, the goal is to create different versions of the image such that the label is the same but the image itself is a bit different. In general, there are many techniques to do that such as cropping, cutout, rotation and others. In this competition, many images have some amount of hair on them, so adding artificial hair is a suitable augmentation for our context that is a bit similar to cutout.\n\nYou are right that adding hair complicates the classification task. But it also motivates a model to learn patterns in the areas not covered by hair. During training, we are learning from the same image multiple times, adding differnt artificial hair in different places on each epoch. This could help to fight overfitting and eventually lead to better performance.", "948725": "Thank you for a detailed response!", "949621": "True, I can confirm. The idea seems good but removing hair lowered my LB Score significantly. Its better to stick with adding hair.", "950154": "okay, thank you!" }, "source": "meta" }