My solution is combination of models with higher input resolution (896*896) and get a bronze medal.
\nHowever, I saw that the 1th solution is the combination of so many small models with relatively low input resolution (https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633).
I'm curious that higher input resolution, more models or larger models, which is usually the first choice to try?
", "messages": [ { "id": "1241918", "postDate": "03/17/2021 09:22:22", "content": "My solution is combination of models with higher input resolution (896*896) and get a bronze medal.
\nHowever, I saw that the 1th solution is the combination of so many small models with relatively low input resolution (https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633).
I'm curious that higher input resolution, more models or larger models, which is usually the first choice to try?
", "rawMarkdown": "My solution is combination of models with higher input resolution (896*896) and get a bronze medal. \nHowever, I saw that the 1th solution is the combination of so many small models with relatively low input resolution (https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633).\n\nI'm curious that higher input resolution, more models or larger models, which is usually the first choice to try?", "votes": null }, { "id": "1245765", "postDate": "03/20/2021 05:54:12", "content": "Interesting question, Dylan. My two cents is that this competition placed heavy emphasis on where the catheter was positioned, not what it looks like. For this reason, higher resolution and large models seemed more important than in many other image classification tasks. Ensembling was necessary to squeeze the last bit out, but for what it's worth seemed to provide less benefit than in most \"what it looks like\" image classification models.
", "rawMarkdown": "Interesting question, Dylan. My two cents is that this competition placed heavy emphasis on where the catheter was positioned, not what it looks like. For this reason, higher resolution and large models seemed more important than in many other image classification tasks. Ensembling was necessary to squeeze the last bit out, but for what it's worth seemed to provide less benefit than in most \"what it looks like\" image classification models.", "votes": null } ], "comments": [ { "id": 1245765, "author_name": "socated", "author_url": "", "post_date": "03/20/2021 05:54:12", "content": "Interesting question, Dylan. My two cents is that this competition placed heavy emphasis on where the catheter was positioned, not what it looks like. For this reason, higher resolution and large models seemed more important than in many other image classification tasks. Ensembling was necessary to squeeze the last bit out, but for what it's worth seemed to provide less benefit than in most \"what it looks like\" image classification models.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1241918": "My solution is combination of models with higher input resolution (896*896) and get a bronze medal. \nHowever, I saw that the 1th solution is the combination of so many small models with relatively low input resolution (https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633).\n\nI'm curious that higher input resolution, more models or larger models, which is usually the first choice to try?", "1245765": "Interesting question, Dylan. My two cents is that this competition placed heavy emphasis on where the catheter was positioned, not what it looks like. For this reason, higher resolution and large models seemed more important than in many other image classification tasks. Ensembling was necessary to squeeze the last bit out, but for what it's worth seemed to provide less benefit than in most \"what it looks like\" image classification models." }, "source": "meta" }