![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14670471%2F07bd65c280338cf287ff972f40c6344c%2FScreenshot%202026-01-01%20192402.png?generation=1767619715309446&alt=media\")
As the title says, I wanna know what do experienced people do during this mid-end phase of the competitions. I currently am in a position, where I understand the competition, understand what the goal is, have a baseline, have tried a few approaches , tried many tweaks, many of which worked while many didn't. It's starting to feel a bit boring because any approach which I try which is not standard, or is a bit unconventional, doesn't work for most part. \nWhat do you guys do in such situations, or what you guys are doing currently? Trying new architectures? trying better post processing methods? Trying newer losses? or simply improving your baseline via higher epochs, higher resolution, more augmentations etc.
", "messages": [ { "id": "3386427", "postDate": "01/05/2026 09:32:50", "content": "As the title says, I wanna know what do experienced people do during this mid-end phase of the competitions. I currently am in a position, where I understand the competition, understand what the goal is, have a baseline, have tried a few approaches , tried many tweaks, many of which worked while many didn't. It's starting to feel a bit boring because any approach which I try which is not standard, or is a bit unconventional, doesn't work for most part. \nWhat do you guys do in such situations, or what you guys are doing currently? Trying new architectures? trying better post processing methods? Trying newer losses? or simply improving your baseline via higher epochs, higher resolution, more augmentations etc.
", "rawMarkdown": "As the title says, I wanna know what do experienced people do during this mid-end phase of the competitions. I currently am in a position, where I understand the competition, understand what the goal is, have a baseline, have tried a few approaches , tried many tweaks, many of which worked while many didn't. It's starting to feel a bit boring because any approach which I try which is not standard, or is a bit unconventional, doesn't work for most part. \nWhat do you guys do in such situations, or what you guys are doing currently? Trying new architectures? trying better post processing methods? Trying newer losses? or simply improving your baseline via higher epochs, higher resolution, more augmentations etc.", "votes": null }, { "id": "3386543", "postDate": "01/05/2026 13:36:25", "content": "I'm a bit new to vision competitions but I assume iterating is similar to LLM competitions. I'm still tryna figure the holes issue
\nI was initially thrown off by the new dataset where my model predict sheets with many holes in it compared to older data where touching sheets was the major problem.
\n
If u already worked on either fixing sheet filling and separating touching sheets (assuming these are common problems), u can try looking at other architectures (I was planning to have a look at the nnUNet architecture as the host was suggesting a 0.59 lb based on training nnUNet and augmentations which is quite impressive. I'm curious to see what makes his architecture superior to the U-Net I'm using right now? is it just longer training with parameters optimally thresholded based on compute or something more)
", "rawMarkdown": "I'm a bit new to vision competitions but I assume iterating is similar to LLM competitions. I'm still tryna figure the holes issue\n\nI was initially thrown off by the new dataset where my model predict sheets with many holes in it compared to older data where touching sheets was the major problem.\n\n\n\nIf u already worked on either fixing sheet filling and separating touching sheets (assuming these are common problems), u can try looking at other architectures (I was planning to have a look at the nnUNet architecture as the host was suggesting a 0.59 lb based on training nnUNet and augmentations which is quite impressive. I'm curious to see what makes his architecture superior to the U-Net I'm using right now? is it just longer training with parameters optimally thresholded based on compute or something more)", "votes": null }, { "id": "3386553", "postDate": "01/05/2026 13:54:55", "content": "Trying new architectures, post-processing methods, and updated loss functions—you're absolutely right.\nYou just focus on doing what you can and leave the rest up to God.
", "rawMarkdown": "Trying new architectures, post-processing methods, and updated loss functions—you're absolutely right.\nYou just focus on doing what you can and leave the rest up to God.", "votes": null }, { "id": "3386707", "postDate": "01/05/2026 18:05:13", "content": "Are you doing inference via patching or downsampling -> inference -> upsampling?
", "rawMarkdown": "Are you doing inference via patching or downsampling -> inference -> upsampling?", "votes": null }, { "id": "3386928", "postDate": "01/06/2026 05:01:16", "content": "From the perspective of Kaggle's workload, I think it's a meaningful way to spend your time to start writing the solution that will be published at the end of the competition.
\nMany participants publish beautiful solutions immediately after the competition closes, and participants are excited to see the top solutions.
", "rawMarkdown": "From the perspective of Kaggle's workload, I think it's a meaningful way to spend your time to start writing the solution that will be published at the end of the competition. \n\nMany participants publish beautiful solutions immediately after the competition closes, and participants are excited to see the top solutions.", "votes": null }, { "id": "3387071", "postDate": "01/06/2026 10:23:45", "content": "just simple inference/ semantic segmentation on U-Net, not patching/instance segmentation
", "rawMarkdown": "just simple inference/ semantic segmentation on U-Net, not patching/instance segmentation", "votes": null }, { "id": "3387074", "postDate": "01/06/2026 10:43:08", "content": "Interesting, will start making a writeup before the competition winds up.
", "rawMarkdown": "Interesting, will start making a writeup before the competition winds up.", "votes": null }, { "id": "3387077", "postDate": "01/06/2026 10:46:07", "content": "I think he meant if you're using some smaller patches like 128x128x128 or downsampling them and then passing that through the network.
", "rawMarkdown": "I think he meant if you're using some smaller patches like 128x128x128 or downsampling them and then passing that through the network.", "votes": null }, { "id": "3387122", "postDate": "01/06/2026 12:23:14", "content": "Ah okay, yes I'm using inference on 160x160x160 patches
", "rawMarkdown": "Ah okay, yes I'm using inference on 160x160x160 patches", "votes": null }, { "id": "3387338", "postDate": "01/06/2026 18:50:12", "content": "Patching refers to the use of sliding window inference to predict patches and stitch them together ig.\nSince not all input images are of the same dims or due to vram constraints instead of infering on the entire volume we do the above.\nHope this helps
", "rawMarkdown": "Patching refers to the use of sliding window inference to predict patches and stitch them together ig.\nSince not all input images are of the same dims or due to vram constraints instead of infering on the entire volume we do the above.\nHope this helps", "votes": null } ], "comments": [ { "id": 3386543, "author_name": "madarshbb", "author_url": "", "post_date": "01/05/2026 13:36:25", "content": "I'm a bit new to vision competitions but I assume iterating is similar to LLM competitions. I'm still tryna figure the holes issue
\nI was initially thrown off by the new dataset where my model predict sheets with many holes in it compared to older data where touching sheets was the major problem.
\n
If u already worked on either fixing sheet filling and separating touching sheets (assuming these are common problems), u can try looking at other architectures (I was planning to have a look at the nnUNet architecture as the host was suggesting a 0.59 lb based on training nnUNet and augmentations which is quite impressive. I'm curious to see what makes his architecture superior to the U-Net I'm using right now? is it just longer training with parameters optimally thresholded based on compute or something more)
", "votes": null, "replies": [ { "id": 3386707, "author_name": "wzyfromhust", "author_url": "", "post_date": "01/05/2026 18:05:13", "content": "Are you doing inference via patching or downsampling -> inference -> upsampling?
", "votes": null, "replies": [ { "id": 3387071, "author_name": "madarshbb", "author_url": "", "post_date": "01/06/2026 10:23:45", "content": "just simple inference/ semantic segmentation on U-Net, not patching/instance segmentation
", "votes": null, "replies": [ { "id": 3387077, "author_name": "choudharymanas", "author_url": "", "post_date": "01/06/2026 10:46:07", "content": "I think he meant if you're using some smaller patches like 128x128x128 or downsampling them and then passing that through the network.
", "votes": null, "replies": [ { "id": 3387122, "author_name": "madarshbb", "author_url": "", "post_date": "01/06/2026 12:23:14", "content": "Ah okay, yes I'm using inference on 160x160x160 patches
", "votes": null, "replies": [ { "id": 3387338, "author_name": "arjunashokbhandary", "author_url": "", "post_date": "01/06/2026 18:50:12", "content": "Patching refers to the use of sliding window inference to predict patches and stitch them together ig.\nSince not all input images are of the same dims or due to vram constraints instead of infering on the entire volume we do the above.\nHope this helps
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3386553, "author_name": "ggayoayogg", "author_url": "", "post_date": "01/05/2026 13:54:55", "content": "Trying new architectures, post-processing methods, and updated loss functions—you're absolutely right.\nYou just focus on doing what you can and leave the rest up to God.
", "votes": null, "replies": [] }, { "id": 3386928, "author_name": "hideyukizushi", "author_url": "", "post_date": "01/06/2026 05:01:16", "content": "From the perspective of Kaggle's workload, I think it's a meaningful way to spend your time to start writing the solution that will be published at the end of the competition.
\nMany participants publish beautiful solutions immediately after the competition closes, and participants are excited to see the top solutions.
", "votes": null, "replies": [ { "id": 3387074, "author_name": "choudharymanas", "author_url": "", "post_date": "01/06/2026 10:43:08", "content": "Interesting, will start making a writeup before the competition winds up.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3386427": "As the title says, I wanna know what do experienced people do during this mid-end phase of the competitions. I currently am in a position, where I understand the competition, understand what the goal is, have a baseline, have tried a few approaches , tried many tweaks, many of which worked while many didn't. It's starting to feel a bit boring because any approach which I try which is not standard, or is a bit unconventional, doesn't work for most part. \nWhat do you guys do in such situations, or what you guys are doing currently? Trying new architectures? trying better post processing methods? Trying newer losses? or simply improving your baseline via higher epochs, higher resolution, more augmentations etc.", "3386543": "I'm a bit new to vision competitions but I assume iterating is similar to LLM competitions. I'm still tryna figure the holes issue\n\nI was initially thrown off by the new dataset where my model predict sheets with many holes in it compared to older data where touching sheets was the major problem.\n\n\n\nIf u already worked on either fixing sheet filling and separating touching sheets (assuming these are common problems), u can try looking at other architectures (I was planning to have a look at the nnUNet architecture as the host was suggesting a 0.59 lb based on training nnUNet and augmentations which is quite impressive. I'm curious to see what makes his architecture superior to the U-Net I'm using right now? is it just longer training with parameters optimally thresholded based on compute or something more)", "3386553": "Trying new architectures, post-processing methods, and updated loss functions—you're absolutely right.\nYou just focus on doing what you can and leave the rest up to God.", "3386707": "Are you doing inference via patching or downsampling -> inference -> upsampling?", "3386928": "From the perspective of Kaggle's workload, I think it's a meaningful way to spend your time to start writing the solution that will be published at the end of the competition. \n\nMany participants publish beautiful solutions immediately after the competition closes, and participants are excited to see the top solutions.", "3387071": "just simple inference/ semantic segmentation on U-Net, not patching/instance segmentation", "3387074": "Interesting, will start making a writeup before the competition winds up.", "3387077": "I think he meant if you're using some smaller patches like 128x128x128 or downsampling them and then passing that through the network.", "3387122": "Ah okay, yes I'm using inference on 160x160x160 patches", "3387338": "Patching refers to the use of sliding window inference to predict patches and stitch them together ig.\nSince not all input images are of the same dims or due to vram constraints instead of infering on the entire volume we do the above.\nHope this helps" }, "source": "meta" }