Just an small comment on the contest of the images, it seems an alpha channel was added, in order to display only the image of the jaguar, however, the RG and B channels still contain the whole photo.
\nI don't know if this was intended or not but is good to know to avoid using the background to identify each animal, as it could harm the training and inference capabilities of your models, but it could also allow to cheat in this competition based on the background of the photo, as the overview indicates:
\n\n\nSpurious Correlations: Models can \"cheat\" by learning the background or the riverbank where a jaguar is seen, rather than the animal's unique patterns.
\n
if I'm misunderstanding the purpose of this information the rest of the channels I apologize for the inconvenience.
", "messages": [ { "id": "3394954", "postDate": "01/22/2026 00:04:18", "content": "Just an small comment on the contest of the images, it seems an alpha channel was added, in order to display only the image of the jaguar, however, the RG and B channels still contain the whole photo.
\nI don't know if this was intended or not but is good to know to avoid using the background to identify each animal, as it could harm the training and inference capabilities of your models, but it could also allow to cheat in this competition based on the background of the photo, as the overview indicates:
\n\n\nSpurious Correlations: Models can \"cheat\" by learning the background or the riverbank where a jaguar is seen, rather than the animal's unique patterns.
\n
if I'm misunderstanding the purpose of this information the rest of the channels I apologize for the inconvenience.
", "rawMarkdown": "Just an small comment on the contest of the images, it seems an alpha channel was added, in order to display only the image of the jaguar, however, the RG and B channels still contain the whole photo.\n\nI don't know if this was intended or not but is good to know to avoid using the background to identify each animal, as it could harm the training and inference capabilities of your models, but it could also allow to cheat in this competition based on the background of the photo, as the overview indicates:\n\n>Spurious Correlations: Models can \"cheat\" by learning the background or the riverbank where a jaguar is seen, rather than the animal's unique patterns.\n\nif I'm misunderstanding the purpose of this information the rest of the channels I apologize for the inconvenience.", "votes": null }, { "id": "3395166", "postDate": "01/22/2026 11:34:02", "content": "Great question!
\nI encourage you to use the included alpha masks or to try your own segmentation method while training on the images on the train set. I segmented them using SAM3.
\nThe test will be extended in a couple of weeks to include more images. These images will lack the RGB background. We want to quantify these spurious correlations where the model relies on the background to identify the subject. We are interested in measuring the performance difference.
", "rawMarkdown": "Great question! \n\nI encourage you to use the included alpha masks or to try your own segmentation method while training on the images on the train set. I segmented them using SAM3.\n\nThe test will be extended in a couple of weeks to include more images. These images will lack the RGB background. We want to quantify these spurious correlations where the model relies on the background to identify the subject. We are interested in measuring the performance difference.", "votes": null }, { "id": "3403874", "postDate": "02/09/2026 13:21:20", "content": "good
", "rawMarkdown": "good", "votes": null } ], "comments": [ { "id": 3395166, "author_name": "andandand", "author_url": "", "post_date": "01/22/2026 11:34:02", "content": "Great question!
\nI encourage you to use the included alpha masks or to try your own segmentation method while training on the images on the train set. I segmented them using SAM3.
\nThe test will be extended in a couple of weeks to include more images. These images will lack the RGB background. We want to quantify these spurious correlations where the model relies on the background to identify the subject. We are interested in measuring the performance difference.
", "votes": null, "replies": [ { "id": 3403874, "author_name": "igorpawicki", "author_url": "", "post_date": "02/09/2026 13:21:20", "content": "good
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3394954": "Just an small comment on the contest of the images, it seems an alpha channel was added, in order to display only the image of the jaguar, however, the RG and B channels still contain the whole photo.\n\nI don't know if this was intended or not but is good to know to avoid using the background to identify each animal, as it could harm the training and inference capabilities of your models, but it could also allow to cheat in this competition based on the background of the photo, as the overview indicates:\n\n>Spurious Correlations: Models can \"cheat\" by learning the background or the riverbank where a jaguar is seen, rather than the animal's unique patterns.\n\nif I'm misunderstanding the purpose of this information the rest of the channels I apologize for the inconvenience.", "3395166": "Great question! \n\nI encourage you to use the included alpha masks or to try your own segmentation method while training on the images on the train set. I segmented them using SAM3.\n\nThe test will be extended in a couple of weeks to include more images. These images will lack the RGB background. We want to quantify these spurious correlations where the model relies on the background to identify the subject. We are interested in measuring the performance difference.", "3403874": "good" }, "source": "meta" }