Dear All,
\nWe are now receiving submissions for the innovation track, and if your idea qualifies for this track, we would like to help you develop it outside the competition. Details follow below.
\nMandatory requirement
\nWe would also like to put forward some expectations from our side for the solution to be considered as innovative. Note that the term ‘innovation’ can vary across the community, and the points below are based on what we believe as an innovative solution. Let’s not create an unnecessary debate on this :)
\nExpectations in a innovative solution (not mandatory if you have a very interesting idea)
\nIf you are really excited about research, then the above points should not be hard to comply with. Build an interesting solution, and let us work together to scale it up.
\nHow to proceed for the innovation track?
\nDear All,
\nWe are now receiving submissions for the innovation track, and if your idea qualifies for this track, we would like to help you develop it outside the competition. Details follow below.
\nMandatory requirement
\nWe would also like to put forward some expectations from our side for the solution to be considered as innovative. Note that the term ‘innovation’ can vary across the community, and the points below are based on what we believe as an innovative solution. Let’s not create an unnecessary debate on this :)
\nExpectations in a innovative solution (not mandatory if you have a very interesting idea)
\nIf you are really excited about research, then the above points should not be hard to comply with. Build an interesting solution, and let us work together to scale it up.
\nHow to proceed for the innovation track?
\nHey Deepak
\nI have understood that we can't use pre-processing to remove background but can we use pre-processing to somehow separate/highlight digits to make them differentiate from background so that neurons can easily understand the features of digits (without using any model trained on standard mist dataset, just a simple pre-processing method)?
\nThanks in advance
", "rawMarkdown": "Hey Deepak\n\nI have understood that we can't use pre-processing to remove background but can we use pre-processing to somehow separate/highlight digits to make them differentiate from background so that neurons can easily understand the features of digits (without using any model trained on standard mist dataset, just a simple pre-processing method)?\n\nThanks in advance", "votes": null }, { "id": "1736869", "postDate": "03/27/2022 20:21:53", "content": "Yes, definitely!
", "rawMarkdown": "Yes, definitely!", "votes": null }, { "id": "1736872", "postDate": "03/27/2022 20:26:49", "content": "Okay, thank you
", "rawMarkdown": "Okay, thank you", "votes": null }, { "id": "1755771", "postDate": "04/15/2022 00:54:59", "content": "Hey, I didn't get time to work much into it since I've been working on other competitions but seam-carving seems to be an interesting idea to resize images while keeping important information (like the digits) the same size. It works by deleting seams of unimportant pixels (calculated by an energy function) to resize an image. This can help with reducing the image size in a pipeline while keeping important information for later processing. It might not do well in this competition time but in more nosey imagery (maybe like satellite imagery) it can definitely help.
", "rawMarkdown": "Hey, I didn't get time to work much into it since I've been working on other competitions but seam-carving seems to be an interesting idea to resize images while keeping important information (like the digits) the same size. It works by deleting seams of unimportant pixels (calculated by an energy function) to resize an image. This can help with reducing the image size in a pipeline while keeping important information for later processing. It might not do well in this competition time but in more nosey imagery (maybe like satellite imagery) it can definitely help.", "votes": null }, { "id": "1755985", "postDate": "04/15/2022 06:13:19", "content": "The idea sounds interesting why don't you submit a working notebook about it for the innovation track 😄
", "rawMarkdown": "The idea sounds interesting why don't you submit a working notebook about it for the innovation track 😄", "votes": null } ], "comments": [ { "id": 1736867, "author_name": "malayjoshi21", "author_url": "", "post_date": "03/27/2022 20:18:23", "content": "Hey Deepak
\nI have understood that we can't use pre-processing to remove background but can we use pre-processing to somehow separate/highlight digits to make them differentiate from background so that neurons can easily understand the features of digits (without using any model trained on standard mist dataset, just a simple pre-processing method)?
\nThanks in advance
", "votes": null, "replies": [ { "id": 1736869, "author_name": "dkgupta90", "author_url": "", "post_date": "03/27/2022 20:21:53", "content": "Yes, definitely!
", "votes": null, "replies": [] }, { "id": 1736872, "author_name": "malayjoshi21", "author_url": "", "post_date": "03/27/2022 20:26:49", "content": "Okay, thank you
", "votes": null, "replies": [] } ] }, { "id": 1755771, "author_name": "outwrest", "author_url": "", "post_date": "04/15/2022 00:54:59", "content": "Hey, I didn't get time to work much into it since I've been working on other competitions but seam-carving seems to be an interesting idea to resize images while keeping important information (like the digits) the same size. It works by deleting seams of unimportant pixels (calculated by an energy function) to resize an image. This can help with reducing the image size in a pipeline while keeping important information for later processing. It might not do well in this competition time but in more nosey imagery (maybe like satellite imagery) it can definitely help.
", "votes": null, "replies": [ { "id": 1755985, "author_name": "ubamba98", "author_url": "", "post_date": "04/15/2022 06:13:19", "content": "The idea sounds interesting why don't you submit a working notebook about it for the innovation track 😄
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1736143": "Dear All, \n\nWe are now receiving submissions for the innovation track, and if your idea qualifies for this track, we would like to help you develop it outside the competition. Details follow below.\n\n**Mandatory requirement**\n1. No pre-processing method used outside the end-to-end deep learning pipeline to remove the background. Reducing noise in a subjective manner is fine.\n\nWe would also like to put forward some expectations from our side for the solution to be considered as innovative. Note that the term ‘innovation’ can vary across the community, and the points below are based on what we believe as an innovative solution. Let’s not create an unnecessary debate on this :)\n\n**Expectations in a innovative solution (not mandatory if you have a very interesting idea)**\n1. Complies with the mandatory requirement outlined above.\n2. Scores above the ‘Innovation-Track-Baseline’ in the leaderboard. The baseline is obtained using a simple Pytorch-based deep learning model and a standard training scheme on a 512X512 dataset [available here](https://www.kaggle.com/datasets/beinggakash/resized-ultramnist-512-akash). Starter code related to this [available here](https://www.kaggle.com/code/surajsharan/ultramnist-starter-notebook).\n3. No use of standard MNIST digits to build a parallel dataset for training any model. We expect this since for real-problems that are similar to UltraMNIST, there is no such base dataset that exists and the code developed for UltraMNIST has to be transferable on real-problems.\n4. You do not need to run the first idea on a very powerful computer at the maximum resolution. Develop the idea on 512X512 and we will help you beyond that if the idea is interesting.\n\nIf you are really excited about research, then the above points should not be hard to comply with. Build an interesting solution, and let us work together to scale it up.\n\n**How to proceed for the innovation track?**\n1. Send us an email at deepak.gupta@iitism.ac.in with your submission name on the leaderboard, some details of your idea and if possible a working notebook for us to test (optional for now). We will maintain full privacy of your idea during the course of the competition.\n2. If your idea intrigues us, we will connect with you to discuss more. We will help you scale up your idea on bigger GPUs to see how well it tackles the UltraMNIST classification issue. This will be done outside the competition with details of the results revealed only after the competition ends. \n3. After the competition ends, we will select the top 3 (or more) innovative solutions and help them improve the idea even further if needed. Based on the final results, we will invite you to co-author the joint paper related to your idea.\n4. We then wait for the reviewers to provide their feedback on our submission :)", "1736867": "Hey Deepak\n\nI have understood that we can't use pre-processing to remove background but can we use pre-processing to somehow separate/highlight digits to make them differentiate from background so that neurons can easily understand the features of digits (without using any model trained on standard mist dataset, just a simple pre-processing method)?\n\nThanks in advance", "1736869": "Yes, definitely!", "1736872": "Okay, thank you", "1755771": "Hey, I didn't get time to work much into it since I've been working on other competitions but seam-carving seems to be an interesting idea to resize images while keeping important information (like the digits) the same size. It works by deleting seams of unimportant pixels (calculated by an energy function) to resize an image. This can help with reducing the image size in a pipeline while keeping important information for later processing. It might not do well in this competition time but in more nosey imagery (maybe like satellite imagery) it can definitely help.", "1755985": "The idea sounds interesting why don't you submit a working notebook about it for the innovation track 😄" }, "source": "meta" }