I have generated synthetic images using Falcon Cloud but now I want the images to be added to the dataset on which I will be training permanently, so that I don't have to reupload it everytime and also that the new data is visible to my teammates. Can I create a new dataset as a mix of the original and my synthetic one, and use that instead of the original dataset? Will the competition allow it?
", "messages": [ { "id": "3252552", "postDate": "07/23/2025 01:38:12", "content": "I have generated synthetic images using Falcon Cloud but now I want the images to be added to the dataset on which I will be training permanently, so that I don't have to reupload it everytime and also that the new data is visible to my teammates. Can I create a new dataset as a mix of the original and my synthetic one, and use that instead of the original dataset? Will the competition allow it?
", "rawMarkdown": "I have generated synthetic images using Falcon Cloud but now I want the images to be added to the dataset on which I will be training permanently, so that I don't have to reupload it everytime and also that the new data is visible to my teammates. Can I create a new dataset as a mix of the original and my synthetic one, and use that instead of the original dataset? Will the competition allow it?", "votes": null }, { "id": "3252739", "postDate": "07/23/2025 10:02:40", "content": "If you are using 3LC for YOLO training, you can refer to Points 8 and 9 under Workflows/Train models and analyze training runs with 3LC in Overview. Basically, you may want to create a new 3LC Table with the newly added data samples and then join it with your previous dataset (Table). The joined 3LC Table will contain both previous data and newly added data, and you can train on this joined Table (joined dataset).
", "rawMarkdown": "If you are using 3LC for YOLO training, you can refer to Points 8 and 9 under Workflows/Train models and analyze training runs with 3LC in Overview. Basically, you may want to create a new 3LC Table with the newly added data samples and then join it with your previous dataset (Table). The joined 3LC Table will contain both previous data and newly added data, and you can train on this joined Table (joined dataset).", "votes": null }, { "id": "3252858", "postDate": "07/23/2025 14:54:54", "content": "Yes! We encourage it. As long as you don't add ANY new real world images to the val folder (or train), you can add as many synthetic images to both the train and val as you'd like.
\nThank you for checking.
I just realized that the dataset contains some real-world images, which I accidentally used to train the model. This explains the high score.
", "rawMarkdown": "I just realized that the dataset contains some real-world images, which I accidentally used to train the model. This explains the high score.", "votes": null }, { "id": "3253045", "postDate": "07/24/2025 00:44:11", "content": "@rebekahduality correct me if I'm mistaken. There are 77 real-world images in the dataset and should only be used as validation.
", "rawMarkdown": "rebekahduality correct me if I'm mistaken. There are 77 real-world images in the dataset and should only be used as validation.", "votes": null }, { "id": "3253147", "postDate": "07/24/2025 05:03:08", "content": "Adding real world images will boost the score unlike the Synthetic images, However the competition requires us to only train on falcon synthetic data.
", "rawMarkdown": "Adding real world images will boost the score unlike the Synthetic images, However the competition requires us to only train on falcon synthetic data.", "votes": null }, { "id": "3253425", "postDate": "07/24/2025 16:44:32", "content": "Yes, that is correct. We provided 77 real-world images and annotations for the val folder ONLY, not in the training.
", "rawMarkdown": "Yes, that is correct. We provided 77 real-world images and annotations for the val folder ONLY, not in the training.", "votes": null }, { "id": "3253427", "postDate": "07/24/2025 16:45:35", "content": "Your current top score… is that with real-world images in the training folder? not val? We can remove it and you can re-upload your model trained on only synthetic if this is the case.
", "rawMarkdown": "Your current top score... is that with real-world images in the training folder? not val? We can remove it and you can re-upload your model trained on only synthetic if this is the case.", "votes": null }, { "id": "3253451", "postDate": "07/24/2025 17:27:34", "content": "Yes, my top score was achieved with a model that has been trained on real-world images. Could you remove all my previous scores?
", "rawMarkdown": "Yes, my top score was achieved with a model that has been trained on real-world images. Could you remove all my previous scores?", "votes": null }, { "id": "3253460", "postDate": "07/24/2025 17:41:01", "content": "Thank you for your honesty. We have removed all of your previous scores, so you can submit your most recent score achieved without the real-world images in the training folder. Thank you!
", "rawMarkdown": "Thank you for your honesty. We have removed all of your previous scores, so you can submit your most recent score achieved without the real-world images in the training folder. Thank you!", "votes": null }, { "id": "3253587", "postDate": "07/25/2025 00:49:40", "content": "This is why personally, I kind of think It doesnt make sense to have real images provided in the dataset (even if it is meant for validation only), since we assume no real data can be labeled and validation is part of our training loop (we use it for hyperparam tuning, model selection, early stopping etc). if organizers supplied labeled real-val images, they’d be leaking target-domain info into every decision and completely undermine the sim-to-real challenge… 🤷♂️
\n@rebekahduality
Yes you are right to some extent. However in real world cases some times we do require some pre-requisite info of our target domain to pin-point and focus on exactly a particular scenario or angle, not for all generalized cases.
\nFor something like that a val set of real-world target specific images is a must!
For Using Falcon beyond this Competition, A mixture of synthetic and real world Images yields optimal results. You can benefit from the insights provided by @rishikeshjadhav22
\nhttps://discord.com/channels/1279188064157306912/1279193650064130080/threads/1318062659882451065
", "rawMarkdown": "For Using Falcon beyond this Competition, A mixture of synthetic and real world Images yields optimal results. You can benefit from the insights provided by @rishikeshjadhav22 \n\nhttps://discord.com/channels/1279188064157306912/1279193650064130080/threads/1318062659882451065", "votes": null } ], "comments": [ { "id": 3252739, "author_name": "dianhebluware", "author_url": "", "post_date": "07/23/2025 10:02:40", "content": "If you are using 3LC for YOLO training, you can refer to Points 8 and 9 under Workflows/Train models and analyze training runs with 3LC in Overview. Basically, you may want to create a new 3LC Table with the newly added data samples and then join it with your previous dataset (Table). The joined 3LC Table will contain both previous data and newly added data, and you can train on this joined Table (joined dataset).
", "votes": null, "replies": [] }, { "id": 3252858, "author_name": "rebekahduality", "author_url": "", "post_date": "07/23/2025 14:54:54", "content": "Yes! We encourage it. As long as you don't add ANY new real world images to the val folder (or train), you can add as many synthetic images to both the train and val as you'd like.
\nThank you for checking.
I just realized that the dataset contains some real-world images, which I accidentally used to train the model. This explains the high score.
", "votes": null, "replies": [ { "id": 3253045, "author_name": "younesbenalia", "author_url": "", "post_date": "07/24/2025 00:44:11", "content": "@rebekahduality correct me if I'm mistaken. There are 77 real-world images in the dataset and should only be used as validation.
", "votes": null, "replies": [ { "id": 3253425, "author_name": "rebekahduality", "author_url": "", "post_date": "07/24/2025 16:44:32", "content": "Yes, that is correct. We provided 77 real-world images and annotations for the val folder ONLY, not in the training.
", "votes": null, "replies": [] } ] }, { "id": 3253147, "author_name": "muhammadhaaris27083", "author_url": "", "post_date": "07/24/2025 05:03:08", "content": "Adding real world images will boost the score unlike the Synthetic images, However the competition requires us to only train on falcon synthetic data.
", "votes": null, "replies": [] }, { "id": 3253427, "author_name": "rebekahduality", "author_url": "", "post_date": "07/24/2025 16:45:35", "content": "Your current top score… is that with real-world images in the training folder? not val? We can remove it and you can re-upload your model trained on only synthetic if this is the case.
", "votes": null, "replies": [ { "id": 3253451, "author_name": "younesbenalia", "author_url": "", "post_date": "07/24/2025 17:27:34", "content": "Yes, my top score was achieved with a model that has been trained on real-world images. Could you remove all my previous scores?
", "votes": null, "replies": [ { "id": 3253460, "author_name": "rebekahduality", "author_url": "", "post_date": "07/24/2025 17:41:01", "content": "Thank you for your honesty. We have removed all of your previous scores, so you can submit your most recent score achieved without the real-world images in the training folder. Thank you!
", "votes": null, "replies": [] } ] } ] }, { "id": 3253587, "author_name": "thelastsmilodon", "author_url": "", "post_date": "07/25/2025 00:49:40", "content": "This is why personally, I kind of think It doesnt make sense to have real images provided in the dataset (even if it is meant for validation only), since we assume no real data can be labeled and validation is part of our training loop (we use it for hyperparam tuning, model selection, early stopping etc). if organizers supplied labeled real-val images, they’d be leaking target-domain info into every decision and completely undermine the sim-to-real challenge… 🤷♂️
\n@rebekahduality
Yes you are right to some extent. However in real world cases some times we do require some pre-requisite info of our target domain to pin-point and focus on exactly a particular scenario or angle, not for all generalized cases.
\nFor something like that a val set of real-world target specific images is a must!
For Using Falcon beyond this Competition, A mixture of synthetic and real world Images yields optimal results. You can benefit from the insights provided by @rishikeshjadhav22
\nhttps://discord.com/channels/1279188064157306912/1279193650064130080/threads/1318062659882451065
", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3252552": "I have generated synthetic images using Falcon Cloud but now I want the images to be added to the dataset on which I will be training permanently, so that I don't have to reupload it everytime and also that the new data is visible to my teammates. Can I create a new dataset as a mix of the original and my synthetic one, and use that instead of the original dataset? Will the competition allow it?", "3252739": "If you are using 3LC for YOLO training, you can refer to Points 8 and 9 under Workflows/Train models and analyze training runs with 3LC in Overview. Basically, you may want to create a new 3LC Table with the newly added data samples and then join it with your previous dataset (Table). The joined 3LC Table will contain both previous data and newly added data, and you can train on this joined Table (joined dataset).", "3252858": "Yes! We encourage it. As long as you don't add ANY new real world images to the val folder (or train), you can add as many synthetic images to both the train and val as you'd like.\nThank you for checking.", "3253037": "I just realized that the dataset contains some real-world images, which I accidentally used to train the model. This explains the high score.", "3253045": "rebekahduality correct me if I'm mistaken. There are 77 real-world images in the dataset and should only be used as validation.", "3253147": "Adding real world images will boost the score unlike the Synthetic images, However the competition requires us to only train on falcon synthetic data.", "3253425": "Yes, that is correct. We provided 77 real-world images and annotations for the val folder ONLY, not in the training.", "3253427": "Your current top score... is that with real-world images in the training folder? not val? We can remove it and you can re-upload your model trained on only synthetic if this is the case.", "3253451": "Yes, my top score was achieved with a model that has been trained on real-world images. Could you remove all my previous scores?", "3253460": "Thank you for your honesty. We have removed all of your previous scores, so you can submit your most recent score achieved without the real-world images in the training folder. Thank you!", "3253587": "This is why personally, I kind of think It doesnt make sense to have real images provided in the dataset (even if it is meant for validation only), since we assume no real data can be labeled and validation is part of our training loop (we use it for hyperparam tuning, model selection, early stopping etc). if organizers supplied labeled real-val images, they’d be leaking target-domain info into every decision and completely undermine the sim-to-real challenge… 🤷♂️ \n@rebekahduality", "3253592": "Yes you are right to some extent. However in real world cases some times we do require some pre-requisite info of our target domain to pin-point and focus on exactly a particular scenario or angle, not for all generalized cases.\nFor something like that a val set of real-world target specific images is a must!", "3253594": "For Using Falcon beyond this Competition, A mixture of synthetic and real world Images yields optimal results. You can benefit from the insights provided by @rishikeshjadhav22 \n\nhttps://discord.com/channels/1279188064157306912/1279193650064130080/threads/1318062659882451065" }, "source": "meta" }