I have been investigating augmenting the YOLO data, flipping the data horizontally or vertically., and training on the result. I see a pretty consistent result: the horizontal flips seem to generate good models; but the vertical flips hurt model performance.
\nWhy? Maybe there’s an anisotropy in the data (i.e. the data doesn’t differ, left-to-right; but does differ top-to-bottom (by top-to-bottom, I mean the y axis, not the z axis). This seems a little unlikely to me, given how the samples were prepared given the discussion in the Project preprint
\nBut perhaps the yolo base models are trained with anisotropic data (i.e. data where reversing left-to-right sounds fine, and reversing up to down doesn’t make a lot of sense..) Looking at the datasets in (https://docs.ultralytics.com/datasets/detect/#ultralytics-yolo-format, many of the datasets (e.g. African Wildlife, COCO…) are camera views of natural scenes, (so there’s often a sky, or there are objects which wouldn’t normally be photographed upside down), or images which have a natural ‘vertical’: brain-tumor.
To find this out, I implemented classes to flip yolo data and the labels.csv files. It probably would have been easier to use ultralytics flip-left-right-fliplr, flip-up-down-flipud or albumentations. Here’s my implementation: Simple Yolo Augmentation
", "messages": [ { "id": "3199975", "postDate": "05/11/2025 19:16:03", "content": "I have been investigating augmenting the YOLO data, flipping the data horizontally or vertically., and training on the result. I see a pretty consistent result: the horizontal flips seem to generate good models; but the vertical flips hurt model performance.
\nWhy? Maybe there’s an anisotropy in the data (i.e. the data doesn’t differ, left-to-right; but does differ top-to-bottom (by top-to-bottom, I mean the y axis, not the z axis). This seems a little unlikely to me, given how the samples were prepared given the discussion in the Project preprint
\nBut perhaps the yolo base models are trained with anisotropic data (i.e. data where reversing left-to-right sounds fine, and reversing up to down doesn’t make a lot of sense..) Looking at the datasets in (https://docs.ultralytics.com/datasets/detect/#ultralytics-yolo-format, many of the datasets (e.g. African Wildlife, COCO…) are camera views of natural scenes, (so there’s often a sky, or there are objects which wouldn’t normally be photographed upside down), or images which have a natural ‘vertical’: brain-tumor.
To find this out, I implemented classes to flip yolo data and the labels.csv files. It probably would have been easier to use ultralytics flip-left-right-fliplr, flip-up-down-flipud or albumentations. Here’s my implementation: Simple Yolo Augmentation
", "rawMarkdown": "I have been investigating augmenting the YOLO data, flipping the data horizontally or vertically., and training on the result. I see a pretty consistent result: the horizontal flips seem to generate good models; but the vertical flips hurt model performance.\n\nWhy? Maybe there’s an anisotropy in the data (i.e. the data doesn’t differ, left-to-right; but does differ top-to-bottom (by top-to-bottom, I mean the y axis, not the z axis). This seems a little unlikely to me, given how the samples were prepared given the discussion in the [Project preprint](https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/575769)\nBut perhaps the yolo base models are trained with anisotropic data (i.e. data where reversing left-to-right sounds fine, and reversing up to down doesn’t make a lot of sense..) Looking at the datasets in (https://docs.ultralytics.com/datasets/detect/#ultralytics-yolo-format, many of the datasets (e.g. [African Wildlife](https://docs.ultralytics.com/datasets/detect/african-wildlife/), [COCO](https://docs.ultralytics.com/datasets/detect/coco/#key-features)…) are camera views of natural scenes, (so there’s often a sky, or there are objects which wouldn’t normally be photographed upside down), or images which have a natural ‘vertical’: [brain-tumor](https://docs.ultralytics.com/datasets/detect/brain-tumor/#usage).\n\n- (There are a few exceptions (I.e. databases, e.g [view](https://docs.ultralytics.com/datasets/detect/xview/#sample-data-and-annotations), [carparts-seg], (https://docs.ultralytics.com/datasets/segment/carparts-seg/#sample-data-and-annotations),\n[crack-seg] (https://docs.ultralytics.com/datasets/segment/crack-seg/) where there is no natural vertical orientation. Perhaps it would be helpful to pre-train with these datasets, or find more datasets that are isotropic, left-to-right and up-to-down.)\n\n\nTo find this out, I implemented classes to flip yolo data and the labels.csv files. It probably would have been easier to use ultralytics flip-left-right-fliplr, flip-up-down-flipud or albumentations. Here’s my implementation: [Simple Yolo Augmentation] (https://www.kaggle.com/code/francisganong/simple-yolo-augmentation)", "votes": null }, { "id": "3200128", "postDate": "05/12/2025 05:37:56", "content": "This is very likely a fluke observation. There should be no reason why one type of a flip would work but not the other, unless the bounding box is calculated incorrectly for vertical flips.
", "rawMarkdown": "This is very likely a fluke observation. There should be no reason why one type of a flip would work but not the other, unless the bounding box is calculated incorrectly for vertical flips.", "votes": null }, { "id": "3200485", "postDate": "05/12/2025 16:34:04", "content": "I agree with him.
", "rawMarkdown": "I agree with him.", "votes": null }, { "id": "3201207", "postDate": "05/13/2025 14:34:06", "content": "I also thought the result was quite suprising. I’ve looked at the code pretty carefully; the handling of ‘vertical flip’ and horizontal flip are quite symmetric, so if you see anything in the code that explains the difference, I would appreciate it.
", "rawMarkdown": "I also thought the result was quite suprising. I’ve looked at the code pretty carefully; the handling of ‘vertical flip’ and horizontal flip are quite symmetric, so if you see anything in the code that explains the difference, I would appreciate it.", "votes": null } ], "comments": [ { "id": 3200128, "author_name": "tilii7", "author_url": "", "post_date": "05/12/2025 05:37:56", "content": "This is very likely a fluke observation. There should be no reason why one type of a flip would work but not the other, unless the bounding box is calculated incorrectly for vertical flips.
", "votes": null, "replies": [ { "id": 3200485, "author_name": "junhanzangai", "author_url": "", "post_date": "05/12/2025 16:34:04", "content": "I agree with him.
", "votes": null, "replies": [ { "id": 3201207, "author_name": "francisganong", "author_url": "", "post_date": "05/13/2025 14:34:06", "content": "I also thought the result was quite suprising. I’ve looked at the code pretty carefully; the handling of ‘vertical flip’ and horizontal flip are quite symmetric, so if you see anything in the code that explains the difference, I would appreciate it.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3199975": "I have been investigating augmenting the YOLO data, flipping the data horizontally or vertically., and training on the result. I see a pretty consistent result: the horizontal flips seem to generate good models; but the vertical flips hurt model performance.\n\nWhy? Maybe there’s an anisotropy in the data (i.e. the data doesn’t differ, left-to-right; but does differ top-to-bottom (by top-to-bottom, I mean the y axis, not the z axis). This seems a little unlikely to me, given how the samples were prepared given the discussion in the [Project preprint](https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/575769)\nBut perhaps the yolo base models are trained with anisotropic data (i.e. data where reversing left-to-right sounds fine, and reversing up to down doesn’t make a lot of sense..) Looking at the datasets in (https://docs.ultralytics.com/datasets/detect/#ultralytics-yolo-format, many of the datasets (e.g. [African Wildlife](https://docs.ultralytics.com/datasets/detect/african-wildlife/), [COCO](https://docs.ultralytics.com/datasets/detect/coco/#key-features)…) are camera views of natural scenes, (so there’s often a sky, or there are objects which wouldn’t normally be photographed upside down), or images which have a natural ‘vertical’: [brain-tumor](https://docs.ultralytics.com/datasets/detect/brain-tumor/#usage).\n\n- (There are a few exceptions (I.e. databases, e.g [view](https://docs.ultralytics.com/datasets/detect/xview/#sample-data-and-annotations), [carparts-seg], (https://docs.ultralytics.com/datasets/segment/carparts-seg/#sample-data-and-annotations),\n[crack-seg] (https://docs.ultralytics.com/datasets/segment/crack-seg/) where there is no natural vertical orientation. Perhaps it would be helpful to pre-train with these datasets, or find more datasets that are isotropic, left-to-right and up-to-down.)\n\n\nTo find this out, I implemented classes to flip yolo data and the labels.csv files. It probably would have been easier to use ultralytics flip-left-right-fliplr, flip-up-down-flipud or albumentations. Here’s my implementation: [Simple Yolo Augmentation] (https://www.kaggle.com/code/francisganong/simple-yolo-augmentation)", "3200128": "This is very likely a fluke observation. There should be no reason why one type of a flip would work but not the other, unless the bounding box is calculated incorrectly for vertical flips.", "3200485": "I agree with him.", "3201207": "I also thought the result was quite suprising. I’ve looked at the code pretty carefully; the handling of ‘vertical flip’ and horizontal flip are quite symmetric, so if you see anything in the code that explains the difference, I would appreciate it." }, "source": "meta" }