I have been very interested in the progress of this competition of seeing different architectures from YOLO to DeIT to GNNs to DETR heads on 3D stacks of 2D embeddings. I am curious what successes and failures between different architectures people have been having so far? I am always fascinated by the different solutions and often learn about different architectures and processing techniques through these competition so I am curious from purely a model architecture POV what everyone has noticed so far. Even from 2D to 2.5D to 3D and tiling the input into smaller patches versus processing whole tomo. Thank you for all participants!
\nNOTE: Please only share what is okay being public, do not share your secret sauce :D
", "messages": [ { "id": "3194328", "postDate": "05/05/2025 17:24:05", "content": "I have been very interested in the progress of this competition of seeing different architectures from YOLO to DeIT to GNNs to DETR heads on 3D stacks of 2D embeddings. I am curious what successes and failures between different architectures people have been having so far? I am always fascinated by the different solutions and often learn about different architectures and processing techniques through these competition so I am curious from purely a model architecture POV what everyone has noticed so far. Even from 2D to 2.5D to 3D and tiling the input into smaller patches versus processing whole tomo. Thank you for all participants!
\nNOTE: Please only share what is okay being public, do not share your secret sauce :D
", "rawMarkdown": "I have been very interested in the progress of this competition of seeing different architectures from YOLO to DeIT to GNNs to DETR heads on 3D stacks of 2D embeddings. I am curious what successes and failures between different architectures people have been having so far? I am always fascinated by the different solutions and often learn about different architectures and processing techniques through these competition so I am curious from purely a model architecture POV what everyone has noticed so far. Even from 2D to 2.5D to 3D and tiling the input into smaller patches versus processing whole tomo. Thank you for all participants! \n\nNOTE: Please only share what is okay being public, do not share your secret sauce :D", "votes": null }, { "id": "3194475", "postDate": "05/05/2025 23:47:59", "content": "In my experience the main issue is not modelling, but building a reliable validation scheme. Because all the methods are more or less obvious (except some exotic ones, such as GNN), but building a reliable CV scheme seems to be a real challenge (at least for me).
", "rawMarkdown": "In my experience the main issue is not modelling, but building a reliable validation scheme. Because all the methods are more or less obvious (except some exotic ones, such as GNN), but building a reliable CV scheme seems to be a real challenge (at least for me).", "votes": null }, { "id": "3195286", "postDate": "05/06/2025 20:30:45", "content": "I don't think you understood the post lol. Different modeling techniques lead to different pre and post processing methods. Size of architecture leads to different types of data loads such as lazy loading. Different outputs (seg mask vs bbox) lead to different techniques to get predictions (such as non maximum suppression). 2D vs 2.5D vs 3D all require different data loading schemes. The point was for people to discuss these different modeling methods and what they have learned from each. For example, I have found non maximum suppression when working with heatmap outputs to work better than threshold tuning to help remove noisy predictions.
", "rawMarkdown": "I don't think you understood the post lol. Different modeling techniques lead to different pre and post processing methods. Size of architecture leads to different types of data loads such as lazy loading. Different outputs (seg mask vs bbox) lead to different techniques to get predictions (such as non maximum suppression). 2D vs 2.5D vs 3D all require different data loading schemes. The point was for people to discuss these different modeling methods and what they have learned from each. For example, I have found non maximum suppression when working with heatmap outputs to work better than threshold tuning to help remove noisy predictions.", "votes": null }, { "id": "3195292", "postDate": "05/06/2025 20:37:26", "content": "Indeed, seems like i didn't understand the meaning of the post lmao. Nevermind ;)
", "rawMarkdown": "Indeed, seems like i didn't understand the meaning of the post lmao. Nevermind ;)", "votes": null } ], "comments": [ { "id": 3194475, "author_name": "eikyou", "author_url": "", "post_date": "05/05/2025 23:47:59", "content": "In my experience the main issue is not modelling, but building a reliable validation scheme. Because all the methods are more or less obvious (except some exotic ones, such as GNN), but building a reliable CV scheme seems to be a real challenge (at least for me).
", "votes": null, "replies": [ { "id": 3195286, "author_name": "connorjd", "author_url": "", "post_date": "05/06/2025 20:30:45", "content": "I don't think you understood the post lol. Different modeling techniques lead to different pre and post processing methods. Size of architecture leads to different types of data loads such as lazy loading. Different outputs (seg mask vs bbox) lead to different techniques to get predictions (such as non maximum suppression). 2D vs 2.5D vs 3D all require different data loading schemes. The point was for people to discuss these different modeling methods and what they have learned from each. For example, I have found non maximum suppression when working with heatmap outputs to work better than threshold tuning to help remove noisy predictions.
", "votes": null, "replies": [ { "id": 3195292, "author_name": "eikyou", "author_url": "", "post_date": "05/06/2025 20:37:26", "content": "Indeed, seems like i didn't understand the meaning of the post lmao. Nevermind ;)
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3194328": "I have been very interested in the progress of this competition of seeing different architectures from YOLO to DeIT to GNNs to DETR heads on 3D stacks of 2D embeddings. I am curious what successes and failures between different architectures people have been having so far? I am always fascinated by the different solutions and often learn about different architectures and processing techniques through these competition so I am curious from purely a model architecture POV what everyone has noticed so far. Even from 2D to 2.5D to 3D and tiling the input into smaller patches versus processing whole tomo. Thank you for all participants! \n\nNOTE: Please only share what is okay being public, do not share your secret sauce :D", "3194475": "In my experience the main issue is not modelling, but building a reliable validation scheme. Because all the methods are more or less obvious (except some exotic ones, such as GNN), but building a reliable CV scheme seems to be a real challenge (at least for me).", "3195286": "I don't think you understood the post lol. Different modeling techniques lead to different pre and post processing methods. Size of architecture leads to different types of data loads such as lazy loading. Different outputs (seg mask vs bbox) lead to different techniques to get predictions (such as non maximum suppression). 2D vs 2.5D vs 3D all require different data loading schemes. The point was for people to discuss these different modeling methods and what they have learned from each. For example, I have found non maximum suppression when working with heatmap outputs to work better than threshold tuning to help remove noisy predictions.", "3195292": "Indeed, seems like i didn't understand the meaning of the post lmao. Nevermind ;)" }, "source": "meta" }