![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2605845%2F8c5c94d68becaed426b4325a52ec3359%2FCropped_bengaliAI.png?generation=1577023738570962&alt=media\")
Hi!
\n\nI saw multiple people working on changing the dataset to a faster .feather type in order to shorten load time.
\n\nAs I'm working on low end hardware (i5 with mx150 gpu & 8Gb of RAM on a laptop) trying to reduce the dataset size is something I try to put some effort in so this was an interesting start.
\n\nSo I've created a dataset of 100x100 images, but rather than simply reducing the image size, I cropped them to only keep the character rather than only the white space around them too.\nThis is a bit more tricky than simply using OpenCV2 and create a bounding box, because some characters contain different independent symbols. So this requires making a exterior bounding box to capture the entire character.
\n\nHere is a quick visual of an original image and a cropped & resized one:
\n\n
For more details, I made a notebook. Feel free to use it if you want!
\n\nAnd the dataset can be found here.
", "messages": [ { "id": 700734, "postDate": "2019-12-22T14:12:21.377Z", "content": "Hi!
\n\nI saw multiple people working on changing the dataset to a faster .feather type in order to shorten load time.
\n\nAs I'm working on low end hardware (i5 with mx150 gpu & 8Gb of RAM on a laptop) trying to reduce the dataset size is something I try to put some effort in so this was an interesting start.
\n\nSo I've created a dataset of 100x100 images, but rather than simply reducing the image size, I cropped them to only keep the character rather than only the white space around them too.\nThis is a bit more tricky than simply using OpenCV2 and create a bounding box, because some characters contain different independent symbols. So this requires making a exterior bounding box to capture the entire character.
\n\nHere is a quick visual of an original image and a cropped & resized one:
\n\n
For more details, I made a notebook. Feel free to use it if you want!
\n\nAnd the dataset can be found here.
", "rawMarkdown": "Hi!\n\nI saw multiple people working on changing the dataset to a faster .feather type in order to shorten load time.\n\nAs I'm working on low end hardware (i5 with mx150 gpu & 8Gb of RAM on a laptop) trying to reduce the dataset size is something I try to put some effort in so this was an interesting start.\n\nSo I've created a dataset of 100x100 images, but rather than simply reducing the image size, I cropped them to only keep the character rather than only the white space around them too.\nThis is a bit more tricky than simply using OpenCV2 and create a bounding box, because some characters contain different independent symbols. So this requires making a exterior bounding box to capture the entire character.\n\nHere is a quick visual of an original image and a cropped & resized one:\n\n\n\nFor more details, I made a [notebook](https://www.kaggle.com/maxlenormand/cropping-to-character-resizing-images). Feel free to use it if you want!\n\nAnd the dataset can be found [here](https://www.kaggle.com/maxlenormand/cropped-resized-bengaliai-images#train_data_3.feather).", "votes": 16 } ], "comments": [], "raw_markdown_by_id": { "700734": "Hi!\n\nI saw multiple people working on changing the dataset to a faster .feather type in order to shorten load time.\n\nAs I'm working on low end hardware (i5 with mx150 gpu & 8Gb of RAM on a laptop) trying to reduce the dataset size is something I try to put some effort in so this was an interesting start.\n\nSo I've created a dataset of 100x100 images, but rather than simply reducing the image size, I cropped them to only keep the character rather than only the white space around them too.\nThis is a bit more tricky than simply using OpenCV2 and create a bounding box, because some characters contain different independent symbols. So this requires making a exterior bounding box to capture the entire character.\n\nHere is a quick visual of an original image and a cropped & resized one:\n\n\n\nFor more details, I made a [notebook](https://www.kaggle.com/maxlenormand/cropping-to-character-resizing-images). Feel free to use it if you want!\n\nAnd the dataset can be found [here](https://www.kaggle.com/maxlenormand/cropped-resized-bengaliai-images#train_data_3.feather)." } }