Hi everyone,
\nWe are happy to announce our fourth annual camera trap challenge, focused on helping to automate data processing and analysis in challenging passive-monitoring cameras placed in the wild and used by ecologists to study animal populations and behavior. The challenge is part of the Eighth Workshop on Fine-Grained Visual Categorization (FGVC8) at CVPR 2021.
\nIn our first challenge, iWildCam2018, we focused on binary empty/animal classification. In the second year of our challenge, iWildCam2019, we focused on generalization to an out-of-sample part of the world with a non-identical set of species. Last year in iWildCam2020 we focused on multimodal data fusion as a way to improve generalization to new camera locations across the globe by providing matched multispectral remote sensing data and iNaturalist data.
\nThis year, we are extending iWildCam2020 to an even more challenging task: counting how many of each species is visible across a sequence of images. This is more challenging than strictly detecting and categorizing species, as it requires reasoning and tracking of individuals across sparse temporal samples. We worked with Centaur Labs to label the test data with species counts, but we do not provide any count labels for the competition as the vast majority of camera trap datasets do not have count labels available. We do provide species labels and weakly-labeled detections from the Microsoft AI for Earth MegaDetector, and the iNaturalist and Remote Sensing data from iWildCam2020. We also did some data cleanup from last year's competition, removing or remapping data with ambiguous labels like \"misfire\" and removing images of humans that were originally mislabeled, adding \"sub_location\" ids for camera locations where ecologists placed more than one camera to maintain background consistency across a location, and correcting sequence labels that were previously in error.
\nLooking forward to the interesting and creative solutions that you all come up with. Feel free to ask questions here, or on the github page!
", "messages": [ { "id": "1234053", "postDate": "03/10/2021 22:22:05", "content": "Hi everyone,
\nWe are happy to announce our fourth annual camera trap challenge, focused on helping to automate data processing and analysis in challenging passive-monitoring cameras placed in the wild and used by ecologists to study animal populations and behavior. The challenge is part of the Eighth Workshop on Fine-Grained Visual Categorization (FGVC8) at CVPR 2021.
\nIn our first challenge, iWildCam2018, we focused on binary empty/animal classification. In the second year of our challenge, iWildCam2019, we focused on generalization to an out-of-sample part of the world with a non-identical set of species. Last year in iWildCam2020 we focused on multimodal data fusion as a way to improve generalization to new camera locations across the globe by providing matched multispectral remote sensing data and iNaturalist data.
\nThis year, we are extending iWildCam2020 to an even more challenging task: counting how many of each species is visible across a sequence of images. This is more challenging than strictly detecting and categorizing species, as it requires reasoning and tracking of individuals across sparse temporal samples. We worked with Centaur Labs to label the test data with species counts, but we do not provide any count labels for the competition as the vast majority of camera trap datasets do not have count labels available. We do provide species labels and weakly-labeled detections from the Microsoft AI for Earth MegaDetector, and the iNaturalist and Remote Sensing data from iWildCam2020. We also did some data cleanup from last year's competition, removing or remapping data with ambiguous labels like \"misfire\" and removing images of humans that were originally mislabeled, adding \"sub_location\" ids for camera locations where ecologists placed more than one camera to maintain background consistency across a location, and correcting sequence labels that were previously in error.
\nLooking forward to the interesting and creative solutions that you all come up with. Feel free to ask questions here, or on the github page!
", "rawMarkdown": "Hi everyone,\n\nWe are happy to announce our fourth annual camera trap challenge, focused on helping to automate data processing and analysis in challenging passive-monitoring cameras placed in the wild and used by ecologists to study animal populations and behavior. The challenge is part of the [Eighth Workshop on Fine-Grained Visual Categorization (FGVC8)](https://sites.google.com/corp/view/fgvc8) at [CVPR 2021](http://cvpr2021.thecvf.com/).\n\nIn our first challenge, [iWildCam2018](https://www.kaggle.com/c/iwildcam2018), we focused on binary empty/animal classification. In the second year of our challenge, [iWildCam2019](https://www.kaggle.com/c/iwildcam-2019-fgvc6), we focused on generalization to an out-of-sample part of the world with a non-identical set of species. Last year in [iWildCam2020](https://www.kaggle.com/c/iwildcam-2020-fgvc7) we focused on multimodal data fusion as a way to improve generalization to new camera locations across the globe by providing matched multispectral remote sensing data and iNaturalist data.\n\nThis year, we are extending iWildCam2020 to an even more challenging task: **counting how many of each species is visible across a sequence of images**. This is more challenging than strictly detecting and categorizing species, as it requires reasoning and tracking of individuals across sparse temporal samples. We worked with [Centaur Labs](https://www.centaurlabs.com/) to label the test data with species counts, but we do not provide any count labels for the competition as the vast majority of camera trap datasets do not have count labels available. We do provide species labels and weakly-labeled detections from the [Microsoft AI for Earth MegaDetector](https://github.com/microsoft/CameraTraps/blob/master/megadetector.md), and the iNaturalist and Remote Sensing data from iWildCam2020. We also did some data cleanup from last year's competition, removing or remapping data with ambiguous labels like \"misfire\" and removing images of humans that were originally mislabeled, adding \"sub_location\" ids for camera locations where ecologists placed more than one camera to maintain background consistency across a location, and correcting sequence labels that were previously in error.\n\nLooking forward to the interesting and creative solutions that you all come up with. Feel free to ask questions here, or on the [github page](https://github.com/visipedia/iwildcam_comp)!", "votes": null }, { "id": "1235216", "postDate": "03/12/2021 00:31:31", "content": "Hi! Can the order of the IDs in the submission format be random? Kaggle often gives sample submission CSV file, but not for this competition. I thought there was an order to the IDs in the submission file, but do I need to worry about it?
", "rawMarkdown": "Hi! Can the order of the IDs in the [submission format](https://www.kaggle.com/c/iwildcam2021-fgvc8/overview/evaluation) be random? Kaggle often gives sample submission CSV file, but not for this competition. I thought there was an order to the IDs in the submission file, but do I need to worry about it?", "votes": null }, { "id": "1235218", "postDate": "03/12/2021 00:35:39", "content": "There is a sample submission listed as a benchmark, but you're right that it isn't in with the data! I'll ask the kaggle data scientist to make sure to add it asap!
", "rawMarkdown": "There is a sample submission listed as a benchmark, but you're right that it isn't in with the data! I'll ask the kaggle data scientist to make sure to add it asap!", "votes": null }, { "id": "1235220", "postDate": "03/12/2021 00:40:48", "content": "And yes, the order can be random :)
", "rawMarkdown": "And yes, the order can be random :)", "votes": null }, { "id": "1235856", "postDate": "03/12/2021 14:26:11", "content": "Thank you for your reply!!!
", "rawMarkdown": "Thank you for your reply!!!", "votes": null }, { "id": "1236041", "postDate": "03/12/2021 17:41:55", "content": "Sample submission is now accessible with the other data :)
", "rawMarkdown": "Sample submission is now accessible with the other data :)", "votes": null }, { "id": "1236342", "postDate": "03/13/2021 03:08:53", "content": "Thank you! :)
", "rawMarkdown": "Thank you! :)", "votes": null }, { "id": "1242419", "postDate": "03/17/2021 15:49:30", "content": "Hi, we have to submit 11057 rows csv file, but there are 60.2k test images. Which is our target???
", "rawMarkdown": "Hi, we have to submit 11057 rows csv file, but there are 60.2k test images. Which is our target???", "votes": null }, { "id": "1242904", "postDate": "03/17/2021 22:23:27", "content": "The targets are the counts across sequences of images, and the target IDs are the sequence IDs. You can group the sequences of images together using the seq_id field in the image metadata.
Thank you for your reply! I had misunderstanded. I could create my submission file. Thank you.
", "rawMarkdown": "Thank you for your reply! I had misunderstanded. I could create my submission file. Thank you.", "votes": null }, { "id": "1253253", "postDate": "03/26/2021 14:22:47", "content": "Can you please provide a description of the metadata including all fields in metadata? It is not immediately clear what these fields represent.
\nThe submission instructions say to submit predictions of counts of each species identified for a given sequence id. \"species\" is not to be found in any of the metadata. I'm assuming \"category_id\" is the same as \"species\". This is not immediately clear. It would be helpful if this were made clearer.
", "rawMarkdown": "Can you please provide a description of the metadata including all fields in metadata? It is not immediately clear what these fields represent.\n\nThe submission instructions say to submit predictions of counts of each species identified for a given sequence id. \"species\" is not to be found in any of the metadata. I'm assuming \"category_id\" is the same as \"species\". This is not immediately clear. It would be helpful if this were made clearer.", "votes": null }, { "id": "1253296", "postDate": "03/26/2021 15:13:06", "content": "Ah, I see a description of data given on the github page.
", "rawMarkdown": "Ah, I see a description of data given on the github page.", "votes": null }, { "id": "1255443", "postDate": "03/28/2021 20:05:33", "content": "Thanks! To clarify my understanding of the provided training data: Is it correct that all we have as training labels per sequence are what animals exist in a sequence, on a frame-by-frame basis? So, the training data does not include any information about how many times each animal occurs in a sequence (apart from the helpful megadetector boxes), even though this is what we are predicting?
", "rawMarkdown": "Thanks! To clarify my understanding of the provided training data: Is it correct that all we have as training labels per sequence are what animals exist in a sequence, on a frame-by-frame basis? So, the training data does not include any information about how many times each animal occurs in a sequence (apart from the helpful megadetector boxes), even though this is what we are predicting?", "votes": null }, { "id": "1256097", "postDate": "03/29/2021 14:40:06", "content": "Sorry this wasn't more clear! I will add a pointer to the github in the data description.
", "rawMarkdown": "Sorry this wasn't more clear! I will add a pointer to the github in the data description.", "votes": null }, { "id": "1256099", "postDate": "03/29/2021 14:40:50", "content": "Yes, the challenge doesn't have counts for training data.
", "rawMarkdown": "Yes, the challenge doesn't have counts for training data.", "votes": null }, { "id": "1271774", "postDate": "04/12/2021 22:23:10", "content": "Hi, I cannot figure out the mega-detector bounding boxes provided. Some look like [xmin, xmax, ymin, ymax] but then others have max values higher than the min? So I thought it was [x-avg, y-avg, w, h], but that doesn't look right either.
", "rawMarkdown": "Hi, I cannot figure out the mega-detector bounding boxes provided. Some look like [xmin, xmax, ymin, ymax] but then others have max values higher than the min? So I thought it was [x-avg, y-avg, w, h], but that doesn't look right either.", "votes": null }, { "id": "1271777", "postDate": "04/12/2021 22:27:08", "content": "Hi! We describe the format on the competition github page: https://github.com/visipedia/iwildcam_comp#camera-trap-animal-detection-model
", "rawMarkdown": "Hi! We describe the format on the competition github page: https://github.com/visipedia/iwildcam_comp#camera-trap-animal-detection-model", "votes": null }, { "id": "1271778", "postDate": "04/12/2021 22:27:47", "content": "It's [x, y, w, h] where x, y is the upper left point
", "rawMarkdown": "It's [x, y, w, h] where x, y is the upper left point", "votes": null }, { "id": "1271796", "postDate": "04/12/2021 23:15:14", "content": "Ok, thank you so much for the quick reply. I think if you can tell me if this is correct or not, I will be clear. For some examples of X,Y:
\n0,1 would be bottom left corner
\n0,0 would be top left corner
\n1,0 would be top right corner
\n1,1 would be bottom right corner
", "rawMarkdown": "Ok, thank you so much for the quick reply. I think if you can tell me if this is correct or not, I will be clear. For some examples of X,Y: \n\n0,1 would be bottom left corner\n\n0,0 would be top left corner\n\n1,0 would be top right corner\n\n1,1 would be bottom right corner", "votes": null }, { "id": "1271797", "postDate": "04/12/2021 23:18:00", "content": "Yes, I believe those would be the coordinates of the corners of the image!
", "rawMarkdown": "Yes, I believe those would be the coordinates of the corners of the image!", "votes": null }, { "id": "1282160", "postDate": "04/23/2021 16:20:29", "content": "Hi, there are redundant \"Sequence ID\" values in the test set. Images can be >10:1 with the Sequence ID. I'm confused by how to create the submission file since I have detection results by image files.
", "rawMarkdown": "Hi, there are redundant \"Sequence ID\" values in the test set. Images can be >10:1 with the Sequence ID. I'm confused by how to create the submission file since I have detection results by image files.", "votes": null }, { "id": "1282216", "postDate": "04/23/2021 17:06:36", "content": "You will need to define a heuristic to aggregate your results for each image in the sequence, so that you have one predicted count for the entire sequence. One simple idea (which is what we used in our baseline) is to use the maximum number of boxes in any of the images in the sequence as your count prediction! However, this will result in undercounting in some cases since the animals might be moving across the frame in the sequence. Some sort of tracking would possibly fix that issue!
", "rawMarkdown": "You will need to define a heuristic to aggregate your results for each image in the sequence, so that you have one predicted count for the entire sequence. One simple idea (which is what we used in our baseline) is to use the maximum number of boxes in any of the images in the sequence as your count prediction! However, this will result in undercounting in some cases since the animals might be moving across the frame in the sequence. Some sort of tracking would possibly fix that issue!", "votes": null }, { "id": "1282246", "postDate": "04/23/2021 17:46:24", "content": "Thank You!
", "rawMarkdown": "Thank You!", "votes": null } ], "comments": [ { "id": 1235216, "author_name": "nayuts", "author_url": "", "post_date": "03/12/2021 00:31:31", "content": "Hi! Can the order of the IDs in the submission format be random? Kaggle often gives sample submission CSV file, but not for this competition. I thought there was an order to the IDs in the submission file, but do I need to worry about it?
", "votes": null, "replies": [ { "id": 1235218, "author_name": "sbeery", "author_url": "", "post_date": "03/12/2021 00:35:39", "content": "There is a sample submission listed as a benchmark, but you're right that it isn't in with the data! I'll ask the kaggle data scientist to make sure to add it asap!
", "votes": null, "replies": [] }, { "id": 1235220, "author_name": "sbeery", "author_url": "", "post_date": "03/12/2021 00:40:48", "content": "And yes, the order can be random :)
", "votes": null, "replies": [] }, { "id": 1235856, "author_name": "nayuts", "author_url": "", "post_date": "03/12/2021 14:26:11", "content": "Thank you for your reply!!!
", "votes": null, "replies": [] }, { "id": 1236041, "author_name": "sbeery", "author_url": "", "post_date": "03/12/2021 17:41:55", "content": "Sample submission is now accessible with the other data :)
", "votes": null, "replies": [] }, { "id": 1236342, "author_name": "nayuts", "author_url": "", "post_date": "03/13/2021 03:08:53", "content": "Thank you! :)
", "votes": null, "replies": [] }, { "id": 1242419, "author_name": "nayuts", "author_url": "", "post_date": "03/17/2021 15:49:30", "content": "Hi, we have to submit 11057 rows csv file, but there are 60.2k test images. Which is our target???
", "votes": null, "replies": [] }, { "id": 1242904, "author_name": "sbeery", "author_url": "", "post_date": "03/17/2021 22:23:27", "content": "The targets are the counts across sequences of images, and the target IDs are the sequence IDs. You can group the sequences of images together using the seq_id field in the image metadata.
Thank you for your reply! I had misunderstanded. I could create my submission file. Thank you.
", "votes": null, "replies": [] } ] }, { "id": 1253253, "author_name": "aamster", "author_url": "", "post_date": "03/26/2021 14:22:47", "content": "Can you please provide a description of the metadata including all fields in metadata? It is not immediately clear what these fields represent.
\nThe submission instructions say to submit predictions of counts of each species identified for a given sequence id. \"species\" is not to be found in any of the metadata. I'm assuming \"category_id\" is the same as \"species\". This is not immediately clear. It would be helpful if this were made clearer.
", "votes": null, "replies": [ { "id": 1253296, "author_name": "aamster", "author_url": "", "post_date": "03/26/2021 15:13:06", "content": "Ah, I see a description of data given on the github page.
", "votes": null, "replies": [] }, { "id": 1256097, "author_name": "sbeery", "author_url": "", "post_date": "03/29/2021 14:40:06", "content": "Sorry this wasn't more clear! I will add a pointer to the github in the data description.
", "votes": null, "replies": [] } ] }, { "id": 1255443, "author_name": "raivokoot", "author_url": "", "post_date": "03/28/2021 20:05:33", "content": "Thanks! To clarify my understanding of the provided training data: Is it correct that all we have as training labels per sequence are what animals exist in a sequence, on a frame-by-frame basis? So, the training data does not include any information about how many times each animal occurs in a sequence (apart from the helpful megadetector boxes), even though this is what we are predicting?
", "votes": null, "replies": [ { "id": 1256099, "author_name": "sbeery", "author_url": "", "post_date": "03/29/2021 14:40:50", "content": "Yes, the challenge doesn't have counts for training data.
", "votes": null, "replies": [] } ] }, { "id": 1271774, "author_name": "uncletito727", "author_url": "", "post_date": "04/12/2021 22:23:10", "content": "Hi, I cannot figure out the mega-detector bounding boxes provided. Some look like [xmin, xmax, ymin, ymax] but then others have max values higher than the min? So I thought it was [x-avg, y-avg, w, h], but that doesn't look right either.
", "votes": null, "replies": [ { "id": 1271777, "author_name": "sbeery", "author_url": "", "post_date": "04/12/2021 22:27:08", "content": "Hi! We describe the format on the competition github page: https://github.com/visipedia/iwildcam_comp#camera-trap-animal-detection-model
", "votes": null, "replies": [] }, { "id": 1271778, "author_name": "sbeery", "author_url": "", "post_date": "04/12/2021 22:27:47", "content": "It's [x, y, w, h] where x, y is the upper left point
", "votes": null, "replies": [] }, { "id": 1271796, "author_name": "uncletito727", "author_url": "", "post_date": "04/12/2021 23:15:14", "content": "Ok, thank you so much for the quick reply. I think if you can tell me if this is correct or not, I will be clear. For some examples of X,Y:
\n0,1 would be bottom left corner
\n0,0 would be top left corner
\n1,0 would be top right corner
\n1,1 would be bottom right corner
", "votes": null, "replies": [] }, { "id": 1271797, "author_name": "sbeery", "author_url": "", "post_date": "04/12/2021 23:18:00", "content": "Yes, I believe those would be the coordinates of the corners of the image!
", "votes": null, "replies": [] } ] }, { "id": 1282160, "author_name": "uncletito727", "author_url": "", "post_date": "04/23/2021 16:20:29", "content": "Hi, there are redundant \"Sequence ID\" values in the test set. Images can be >10:1 with the Sequence ID. I'm confused by how to create the submission file since I have detection results by image files.
", "votes": null, "replies": [ { "id": 1282216, "author_name": "sbeery", "author_url": "", "post_date": "04/23/2021 17:06:36", "content": "You will need to define a heuristic to aggregate your results for each image in the sequence, so that you have one predicted count for the entire sequence. One simple idea (which is what we used in our baseline) is to use the maximum number of boxes in any of the images in the sequence as your count prediction! However, this will result in undercounting in some cases since the animals might be moving across the frame in the sequence. Some sort of tracking would possibly fix that issue!
", "votes": null, "replies": [] }, { "id": 1282246, "author_name": "uncletito727", "author_url": "", "post_date": "04/23/2021 17:46:24", "content": "Thank You!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1234053": "Hi everyone,\n\nWe are happy to announce our fourth annual camera trap challenge, focused on helping to automate data processing and analysis in challenging passive-monitoring cameras placed in the wild and used by ecologists to study animal populations and behavior. The challenge is part of the [Eighth Workshop on Fine-Grained Visual Categorization (FGVC8)](https://sites.google.com/corp/view/fgvc8) at [CVPR 2021](http://cvpr2021.thecvf.com/).\n\nIn our first challenge, [iWildCam2018](https://www.kaggle.com/c/iwildcam2018), we focused on binary empty/animal classification. In the second year of our challenge, [iWildCam2019](https://www.kaggle.com/c/iwildcam-2019-fgvc6), we focused on generalization to an out-of-sample part of the world with a non-identical set of species. Last year in [iWildCam2020](https://www.kaggle.com/c/iwildcam-2020-fgvc7) we focused on multimodal data fusion as a way to improve generalization to new camera locations across the globe by providing matched multispectral remote sensing data and iNaturalist data.\n\nThis year, we are extending iWildCam2020 to an even more challenging task: **counting how many of each species is visible across a sequence of images**. This is more challenging than strictly detecting and categorizing species, as it requires reasoning and tracking of individuals across sparse temporal samples. We worked with [Centaur Labs](https://www.centaurlabs.com/) to label the test data with species counts, but we do not provide any count labels for the competition as the vast majority of camera trap datasets do not have count labels available. We do provide species labels and weakly-labeled detections from the [Microsoft AI for Earth MegaDetector](https://github.com/microsoft/CameraTraps/blob/master/megadetector.md), and the iNaturalist and Remote Sensing data from iWildCam2020. We also did some data cleanup from last year's competition, removing or remapping data with ambiguous labels like \"misfire\" and removing images of humans that were originally mislabeled, adding \"sub_location\" ids for camera locations where ecologists placed more than one camera to maintain background consistency across a location, and correcting sequence labels that were previously in error.\n\nLooking forward to the interesting and creative solutions that you all come up with. Feel free to ask questions here, or on the [github page](https://github.com/visipedia/iwildcam_comp)!", "1235216": "Hi! Can the order of the IDs in the [submission format](https://www.kaggle.com/c/iwildcam2021-fgvc8/overview/evaluation) be random? Kaggle often gives sample submission CSV file, but not for this competition. I thought there was an order to the IDs in the submission file, but do I need to worry about it?", "1235218": "There is a sample submission listed as a benchmark, but you're right that it isn't in with the data! I'll ask the kaggle data scientist to make sure to add it asap!", "1235220": "And yes, the order can be random :)", "1235856": "Thank you for your reply!!!", "1236041": "Sample submission is now accessible with the other data :)", "1236342": "Thank you! :)", "1242419": "Hi, we have to submit 11057 rows csv file, but there are 60.2k test images. Which is our target???", "1242904": "The targets are the counts across sequences of images, and the target IDs are the sequence IDs. You can group the sequences of images together using the `seq_id` field in the image metadata.", "1243880": "Thank you for your reply! I had misunderstanded. I could create my submission file. Thank you.", "1253253": "Can you please provide a description of the metadata including all fields in metadata? It is not immediately clear what these fields represent.\n\nThe submission instructions say to submit predictions of counts of each species identified for a given sequence id. \"species\" is not to be found in any of the metadata. I'm assuming \"category_id\" is the same as \"species\". This is not immediately clear. It would be helpful if this were made clearer.", "1253296": "Ah, I see a description of data given on the github page.", "1255443": "Thanks! To clarify my understanding of the provided training data: Is it correct that all we have as training labels per sequence are what animals exist in a sequence, on a frame-by-frame basis? So, the training data does not include any information about how many times each animal occurs in a sequence (apart from the helpful megadetector boxes), even though this is what we are predicting?", "1256097": "Sorry this wasn't more clear! I will add a pointer to the github in the data description.", "1256099": "Yes, the challenge doesn't have counts for training data.", "1271774": "Hi, I cannot figure out the mega-detector bounding boxes provided. Some look like [xmin, xmax, ymin, ymax] but then others have max values higher than the min? So I thought it was [x-avg, y-avg, w, h], but that doesn't look right either.", "1271777": "Hi! We describe the format on the competition github page: https://github.com/visipedia/iwildcam_comp#camera-trap-animal-detection-model", "1271778": "It's [x, y, w, h] where x, y is the upper left point", "1271796": "Ok, thank you so much for the quick reply. I think if you can tell me if this is correct or not, I will be clear. For some examples of X,Y: \n\n0,1 would be bottom left corner\n\n0,0 would be top left corner\n\n1,0 would be top right corner\n\n1,1 would be bottom right corner", "1271797": "Yes, I believe those would be the coordinates of the corners of the image!", "1282160": "Hi, there are redundant \"Sequence ID\" values in the test set. Images can be >10:1 with the Sequence ID. I'm confused by how to create the submission file since I have detection results by image files.", "1282216": "You will need to define a heuristic to aggregate your results for each image in the sequence, so that you have one predicted count for the entire sequence. One simple idea (which is what we used in our baseline) is to use the maximum number of boxes in any of the images in the sequence as your count prediction! However, this will result in undercounting in some cases since the animals might be moving across the frame in the sequence. Some sort of tracking would possibly fix that issue!", "1282246": "Thank You!" }, "source": "meta" }