In the previous edition of this comp, a strategy used by many competitors was that to increase the training images, they used to flip the images, creating new images to train.
\nYou can see 1st placed and 3rd placed winners of the last comp saying it in their solutions
Since last time it was the flukes of a whale and this time its the fins I don't know whether this strategy can be used.
\ncan the organiser comment something on it? @tedcheese
", "messages": [ { "id": "1678047", "postDate": "02/06/2022 08:21:05", "content": "In the previous edition of this comp, a strategy used by many competitors was that to increase the training images, they used to flip the images, creating new images to train.
\nYou can see 1st placed and 3rd placed winners of the last comp saying it in their solutions
Since last time it was the flukes of a whale and this time its the fins I don't know whether this strategy can be used.
\ncan the organiser comment something on it? @tedcheese
", "rawMarkdown": "In the previous edition of this comp, a strategy used by many competitors was that to increase the training images, they used to flip the images, creating new images to train.\nYou can see [1st](https://www.kaggle.com/c/humpback-whale-identification/discussion/82366) placed and [3rd](https://www.kaggle.com/c/humpback-whale-identification/discussion/82484) placed winners of the last comp saying it in their solutions\n\nSince last time it was the flukes of a whale and this time its the fins I don't know whether this strategy can be used.\n\ncan the organiser comment something on it? @tedcheese", "votes": null }, { "id": "1678056", "postDate": "02/06/2022 08:24:59", "content": "Hmm, good question. Dolphin dorsal fin might be symmetrical?🤔
", "rawMarkdown": "Hmm, good question. Dolphin dorsal fin might be symmetrical?🤔", "votes": null }, { "id": "1678122", "postDate": "02/06/2022 09:00:58", "content": "so if they are symmetrical, like flukes, flipping it will give us new training data?
", "rawMarkdown": "so if they are symmetrical, like flukes, flipping it will give us new training data?", "votes": null }, { "id": "1678123", "postDate": "02/06/2022 09:03:32", "content": "No, I think more about the opposite - if they are roughly the same from both sides, it might confuse the model even more
", "rawMarkdown": "No, I think more about the opposite - if they are roughly the same from both sides, it might confuse the model even more", "votes": null }, { "id": "1678547", "postDate": "02/06/2022 16:00:56", "content": "It is my expectation that flipping would indeed confuse the model because in many cases it is the shape of the edge that is used, and for many individuals we have sought to have both left side and right side views. A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.
\nThat said, since some individuals may have only a left side view in the training set and only a right side in the test set or vice versa, perhaps for individuals that don't have both views in the training set, flipping those specific training set images may help with a more complete predictive training set?
", "rawMarkdown": "It is my expectation that flipping would indeed confuse the model because in many cases it is the shape of the edge that is used, and for many individuals we have sought to have both left side and right side views. A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.\n\nThat said, since some individuals may have only a left side view in the training set and only a right side in the test set or vice versa, perhaps for individuals that don't have both views in the training set, flipping those specific training set images may help with a more complete predictive training set?", "votes": null }, { "id": "1678626", "postDate": "02/06/2022 17:09:33", "content": "A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.
\nI see then I think we can't use this trick this time, since like you said flipping would still make researchers classify it as the same whale/dolphin.
\nflipping those specific training set images may help with a more complete predictive training set?
\nyes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.
@tedcheese Thanks for the input. I agree with you! @mrinath I think we should try to verify what works but here's my take!
\nFlukes are symmetric, but only one view (rear side) seems to have been used in prev competition. so they are single-sided (rear side). Hence, say a spot on the right side {.\\/ } of a fluke and a spot on left { \\/.} cant be from the same fluke.
Fins are asymmetric, but two views of the same fin can be seen in the data, hence two-sided. But not all kinds of marks would be visible from two sides.
Case - 1: Flip gives new individual
\nA spot on a side -> /.| when flipped is a spot on the other side of the fin |.\\ <- and as the spot would exist on the other side of the fin.
Case - 2: Flip give the same individual
\nA scar on a rear edge /{ <-, when flipped, is still a scar on the rear edge ->}\\, and will be similar to the view from the other side of the fin, so can be from the same fin.
So, it depends on the type of marking that is used to identify the individual, and it's very likely that both types of scars can be noticed on the fins. Hence not flipping would be better.
", "rawMarkdown": "tedcheese Thanks for the input. I agree with you! @mrinath I think we should try to verify what works but here's my take!\n\nFlukes are symmetric, but only one view (rear side) seems to have been used in prev competition. so they are `single-sided` (rear side). Hence, say a spot on the right side {.\\/ } of a fluke and a spot on left { \\/.} cant be from the same fluke.\n\nFins are asymmetric, but two views of the same fin can be seen in the data, hence `two-sided`. But not all kinds of marks would be visible from two sides. \n\n**Case - 1**: Flip gives new individual\nA spot on a side -> /.| when flipped is a spot on the other side of the fin |.\\ <- and as the spot would exist on the other side of the fin. \n\n**Case - 2**: Flip give the same individual\nA scar on a rear edge /{ <-, when flipped, is still a scar on the rear edge ->}\\, and will be similar to the view from the other side of the fin, so can be from the same fin.\n\nSo, it depends on the type of marking that is used to identify the individual, and it's very likely that both types of scars can be noticed on the fins. Hence not flipping would be better.", "votes": null }, { "id": "1678832", "postDate": "02/06/2022 20:10:26", "content": "wow, @bsridatta thanks for explaining it so thoughtfully!!
", "rawMarkdown": "wow, @bsridatta thanks for explaining it so thoughtfully!!", "votes": null }, { "id": "1679033", "postDate": "02/06/2022 23:12:35", "content": "Welcome :) I was a bit confused too so thought to put it down clearly.
\nYou might be interesting in follow this discussion too
\nhttps://www.kaggle.com/c/happy-whale-and-dolphin/discussion/305607
\n\nyes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.
\n
Only one way to find out 😄
", "rawMarkdown": "> yes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.\n\nOnly one way to find out 😄", "votes": null } ], "comments": [ { "id": 1678056, "author_name": "olegsidorshin", "author_url": "", "post_date": "02/06/2022 08:24:59", "content": "Hmm, good question. Dolphin dorsal fin might be symmetrical?🤔
", "votes": null, "replies": [ { "id": 1678122, "author_name": "mrinath", "author_url": "", "post_date": "02/06/2022 09:00:58", "content": "so if they are symmetrical, like flukes, flipping it will give us new training data?
", "votes": null, "replies": [] }, { "id": 1678123, "author_name": "olegsidorshin", "author_url": "", "post_date": "02/06/2022 09:03:32", "content": "No, I think more about the opposite - if they are roughly the same from both sides, it might confuse the model even more
", "votes": null, "replies": [] }, { "id": 1678547, "author_name": "tedcheese", "author_url": "", "post_date": "02/06/2022 16:00:56", "content": "It is my expectation that flipping would indeed confuse the model because in many cases it is the shape of the edge that is used, and for many individuals we have sought to have both left side and right side views. A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.
\nThat said, since some individuals may have only a left side view in the training set and only a right side in the test set or vice versa, perhaps for individuals that don't have both views in the training set, flipping those specific training set images may help with a more complete predictive training set?
", "votes": null, "replies": [] }, { "id": 1678626, "author_name": "mrinath", "author_url": "", "post_date": "02/06/2022 17:09:33", "content": "A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.
\nI see then I think we can't use this trick this time, since like you said flipping would still make researchers classify it as the same whale/dolphin.
\nflipping those specific training set images may help with a more complete predictive training set?
\nyes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.
@tedcheese Thanks for the input. I agree with you! @mrinath I think we should try to verify what works but here's my take!
\nFlukes are symmetric, but only one view (rear side) seems to have been used in prev competition. so they are single-sided (rear side). Hence, say a spot on the right side {.\\/ } of a fluke and a spot on left { \\/.} cant be from the same fluke.
Fins are asymmetric, but two views of the same fin can be seen in the data, hence two-sided. But not all kinds of marks would be visible from two sides.
Case - 1: Flip gives new individual
\nA spot on a side -> /.| when flipped is a spot on the other side of the fin |.\\ <- and as the spot would exist on the other side of the fin.
Case - 2: Flip give the same individual
\nA scar on a rear edge /{ <-, when flipped, is still a scar on the rear edge ->}\\, and will be similar to the view from the other side of the fin, so can be from the same fin.
So, it depends on the type of marking that is used to identify the individual, and it's very likely that both types of scars can be noticed on the fins. Hence not flipping would be better.
", "votes": null, "replies": [] }, { "id": 1678832, "author_name": "mrinath", "author_url": "", "post_date": "02/06/2022 20:10:26", "content": "wow, @bsridatta thanks for explaining it so thoughtfully!!
", "votes": null, "replies": [] }, { "id": 1679033, "author_name": "bsridatta", "author_url": "", "post_date": "02/06/2022 23:12:35", "content": "Welcome :) I was a bit confused too so thought to put it down clearly.
\nYou might be interesting in follow this discussion too
\nhttps://www.kaggle.com/c/happy-whale-and-dolphin/discussion/305607
\n\nyes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.
\n
Only one way to find out 😄
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1678047": "In the previous edition of this comp, a strategy used by many competitors was that to increase the training images, they used to flip the images, creating new images to train.\nYou can see [1st](https://www.kaggle.com/c/humpback-whale-identification/discussion/82366) placed and [3rd](https://www.kaggle.com/c/humpback-whale-identification/discussion/82484) placed winners of the last comp saying it in their solutions\n\nSince last time it was the flukes of a whale and this time its the fins I don't know whether this strategy can be used.\n\ncan the organiser comment something on it? @tedcheese", "1678056": "Hmm, good question. Dolphin dorsal fin might be symmetrical?🤔", "1678122": "so if they are symmetrical, like flukes, flipping it will give us new training data?", "1678123": "No, I think more about the opposite - if they are roughly the same from both sides, it might confuse the model even more", "1678547": "It is my expectation that flipping would indeed confuse the model because in many cases it is the shape of the edge that is used, and for many individuals we have sought to have both left side and right side views. A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.\n\nThat said, since some individuals may have only a left side view in the training set and only a right side in the test set or vice versa, perhaps for individuals that don't have both views in the training set, flipping those specific training set images may help with a more complete predictive training set?", "1678626": "`A flipped view of an individual will usually not be independent, and will likely be closely correlated with a view of the same individual from the opposite side.`\nI see then I think we **can't** use this trick this time, since like you said flipping would still make researchers classify it as the same whale/dolphin.\n`flipping those specific training set images may help with a more complete predictive training set?`\nyes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.", "1678747": "tedcheese Thanks for the input. I agree with you! @mrinath I think we should try to verify what works but here's my take!\n\nFlukes are symmetric, but only one view (rear side) seems to have been used in prev competition. so they are `single-sided` (rear side). Hence, say a spot on the right side {.\\/ } of a fluke and a spot on left { \\/.} cant be from the same fluke.\n\nFins are asymmetric, but two views of the same fin can be seen in the data, hence `two-sided`. But not all kinds of marks would be visible from two sides. \n\n**Case - 1**: Flip gives new individual\nA spot on a side -> /.| when flipped is a spot on the other side of the fin |.\\ <- and as the spot would exist on the other side of the fin. \n\n**Case - 2**: Flip give the same individual\nA scar on a rear edge /{ <-, when flipped, is still a scar on the rear edge ->}\\, and will be similar to the view from the other side of the fin, so can be from the same fin.\n\nSo, it depends on the type of marking that is used to identify the individual, and it's very likely that both types of scars can be noticed on the fins. Hence not flipping would be better.", "1678832": "wow, @bsridatta thanks for explaining it so thoughtfully!!", "1679033": "Welcome :) I was a bit confused too so thought to put it down clearly. \n\nYou might be interesting in follow this discussion too \nhttps://www.kaggle.com/c/happy-whale-and-dolphin/discussion/305607", "1679338": "> yes, I think it will help, adding a flip augmentation in our training pipeline would definitely help the model.\n\nOnly one way to find out 😄" }, "source": "meta" }