Hi everyone,\nI’m quite new to Kaggle and to plant-phenotyping tasks.\nWhile studying the public notebooks, I noticed that many of the top solutions are using frozen DINOv2 (large or giant) as the feature extractor.\nIt seems to have become the default choice.\nAs a beginner, I’m wondering:\nIs DINOv2 generally considered the strongest backbone right now for this kind of UAV/drone biomass regression competition, or are there other models (ConvNeXt-V2, SigLIP, InternImage, newer DINOv3, CLIP-based, or agriculture-specific pretrained models) that people have tried and found to perform similarly or even better in CV/LB?\nWould love to hear your real experiences and any score comparisons. Thank you!
", "messages": [ { "id": "3352736", "postDate": "11/29/2025 11:32:05", "content": "Hi everyone,\nI’m quite new to Kaggle and to plant-phenotyping tasks.\nWhile studying the public notebooks, I noticed that many of the top solutions are using frozen DINOv2 (large or giant) as the feature extractor.\nIt seems to have become the default choice.\nAs a beginner, I’m wondering:\nIs DINOv2 generally considered the strongest backbone right now for this kind of UAV/drone biomass regression competition, or are there other models (ConvNeXt-V2, SigLIP, InternImage, newer DINOv3, CLIP-based, or agriculture-specific pretrained models) that people have tried and found to perform similarly or even better in CV/LB?\nWould love to hear your real experiences and any score comparisons. Thank you!
", "rawMarkdown": "Hi everyone,\nI’m quite new to Kaggle and to plant-phenotyping tasks.\nWhile studying the public notebooks, I noticed that many of the top solutions are using frozen DINOv2 (large or giant) as the feature extractor.\nIt seems to have become the default choice.\nAs a beginner, I’m wondering:\nIs DINOv2 generally considered the strongest backbone right now for this kind of UAV/drone biomass regression competition, or are there other models (ConvNeXt-V2, SigLIP, InternImage, newer DINOv3, CLIP-based, or agriculture-specific pretrained models) that people have tried and found to perform similarly or even better in CV/LB?\nWould love to hear your real experiences and any score comparisons. Thank you!", "votes": null }, { "id": "3353282", "postDate": "11/29/2025 16:46:58", "content": "Hi @lirentu. I would say it would be unfair to share scores of such models. However, what I can tell you is that there is no 'default choice' here. Some models are simply 'easier' to use compared to the others requiring less-to-no fine tuning. But, once you fine tune these models, you will realize that you can get comparable results with any of them.
\nLet me also tell you that for me, I have seen these models perform quite differently on different folds. The extreme case would be for example 5 models which performs the best on one of the 5 folds. This mainly happens because of the small dataset, distribution shift between train and validation split and also the training that the model had undergone instead of the model itself. Dino series for example went under more self supervised learning compared to supervised learning which might be the reason it performs slightly better as frozen extractor for regression tasks.
", "rawMarkdown": "Hi @lirentu. I would say it would be unfair to share scores of such models. However, what I can tell you is that there is no 'default choice' here. Some models are simply 'easier' to use compared to the others requiring less-to-no fine tuning. But, once you fine tune these models, you will realize that you can get comparable results with any of them. \n\nLet me also tell you that for me, I have seen these models perform quite differently on different folds. The extreme case would be for example 5 models which performs the best on one of the 5 folds. This mainly happens because of the small dataset, distribution shift between train and validation split and also the training that the model had undergone instead of the model itself. Dino series for example went under more self supervised learning compared to supervised learning which might be the reason it performs slightly better as frozen extractor for regression tasks.", "votes": null }, { "id": "3353493", "postDate": "11/29/2025 19:00:08", "content": "Our best model is not DINOv2. It also depends on the strategy you’re using.
", "rawMarkdown": "Our best model is not DINOv2. It also depends on the strategy you’re using.", "votes": null } ], "comments": [ { "id": 3353282, "author_name": "komilparmar", "author_url": "", "post_date": "11/29/2025 16:46:58", "content": "Hi @lirentu. I would say it would be unfair to share scores of such models. However, what I can tell you is that there is no 'default choice' here. Some models are simply 'easier' to use compared to the others requiring less-to-no fine tuning. But, once you fine tune these models, you will realize that you can get comparable results with any of them.
\nLet me also tell you that for me, I have seen these models perform quite differently on different folds. The extreme case would be for example 5 models which performs the best on one of the 5 folds. This mainly happens because of the small dataset, distribution shift between train and validation split and also the training that the model had undergone instead of the model itself. Dino series for example went under more self supervised learning compared to supervised learning which might be the reason it performs slightly better as frozen extractor for regression tasks.
", "votes": null, "replies": [] }, { "id": 3353493, "author_name": "yeoyunsianggeremie", "author_url": "", "post_date": "11/29/2025 19:00:08", "content": "Our best model is not DINOv2. It also depends on the strategy you’re using.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3352736": "Hi everyone,\nI’m quite new to Kaggle and to plant-phenotyping tasks.\nWhile studying the public notebooks, I noticed that many of the top solutions are using frozen DINOv2 (large or giant) as the feature extractor.\nIt seems to have become the default choice.\nAs a beginner, I’m wondering:\nIs DINOv2 generally considered the strongest backbone right now for this kind of UAV/drone biomass regression competition, or are there other models (ConvNeXt-V2, SigLIP, InternImage, newer DINOv3, CLIP-based, or agriculture-specific pretrained models) that people have tried and found to perform similarly or even better in CV/LB?\nWould love to hear your real experiences and any score comparisons. Thank you!", "3353282": "Hi @lirentu. I would say it would be unfair to share scores of such models. However, what I can tell you is that there is no 'default choice' here. Some models are simply 'easier' to use compared to the others requiring less-to-no fine tuning. But, once you fine tune these models, you will realize that you can get comparable results with any of them. \n\nLet me also tell you that for me, I have seen these models perform quite differently on different folds. The extreme case would be for example 5 models which performs the best on one of the 5 folds. This mainly happens because of the small dataset, distribution shift between train and validation split and also the training that the model had undergone instead of the model itself. Dino series for example went under more self supervised learning compared to supervised learning which might be the reason it performs slightly better as frozen extractor for regression tasks.", "3353493": "Our best model is not DINOv2. It also depends on the strategy you’re using." }, "source": "meta" }