![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1739631%2F8817818d0fd3f061beb56ca9c9a545be%2F22.PNG?generation=1590351230256178&alt=media\")
1,2,3 use albumentations
", "messages": [ { "id": 852143, "postDate": "2020-05-18T07:48:03.330Z", "content": "1,2,3 use albumentations
", "rawMarkdown": "1. random vertical-flip and horizontal-flip\n2. random BrightnessContrast\t\n3. random HueSaturationValue\n4. Multiscale Test\n\n1,2,3 use [albumentations](https://github.com/albumentations-team/albumentations)\n", "votes": 18 }, { "id": 852626, "postDate": "2020-05-18T15:08:44.887Z", "content": "@spytensor Thanks for sharing. I had one question: are you creating new larger dataset with this augmentation or just augmenting on the fly while training? I've tried some of these and my score decresed instead of increasing. I'm using original image size with 25 epochs of FasterRCNN.
", "rawMarkdown": "@spytensor Thanks for sharing. I had one question: are you creating new larger dataset with this augmentation or just augmenting on the fly while training? I've tried some of these and my score decresed instead of increasing. I'm using original image size with 25 epochs of FasterRCNN. ", "votes": 1, "replies": [ { "id": 852971, "postDate": "2020-05-18T21:07:21.857Z", "content": "I am also curious about this.🤔
", "rawMarkdown": "I am also curious about this.🤔 " }, { "id": 853142, "postDate": "2020-05-19T01:57:37.517Z", "content": "Randomly do augments while training models.
", "rawMarkdown": "Randomly do augments while training models.", "votes": 3 }, { "id": 853206, "postDate": "2020-05-19T03:06:46.323Z", "content": "Cool! Thanks. @spytensor
", "rawMarkdown": "Cool! Thanks. @spytensor " } ] }, { "id": 852550, "postDate": "2020-05-18T14:25:31.547Z", "content": "Thanks for sharing! adding to your list, \nrandom sized BBoxSafeCrop worked for me, also using albumentations
", "rawMarkdown": "Thanks for sharing! adding to your list, \nrandom sized BBoxSafeCrop worked for me, also using albumentations ", "votes": 1, "replies": [ { "id": 853144, "postDate": "2020-05-19T01:59:19.773Z", "content": "👍 ！Thanks ,I'll try it later.
", "rawMarkdown": "👍 ！Thanks ,I'll try it later." }, { "id": 854290, "postDate": "2020-05-20T00:01:45.697Z", "content": "Hello! I'm testing RandomSizedBBoxSafeCrop, but sometimes it brings error which tells me the bbox number is 0. So hao to set he parameters of this trick especially erosion_rate?
", "rawMarkdown": "Hello! I'm testing RandomSizedBBoxSafeCrop, but sometimes it brings error which tells me the bbox number is 0. So hao to set he parameters of this trick especially erosion_rate?" }, { "id": 855371, "postDate": "2020-05-20T21:37:02.343Z", "content": "", "rawMarkdown": "Refer [here](https://www.kaggle.com/c/global-wheat-detection/discussion/152595) or [here](https://www.kaggle.com/c/global-wheat-detection/discussion/152423).", "votes": 1 }, { "id": 857066, "postDate": "2020-05-22T09:20:57.860Z", "content": "Thanks!
", "rawMarkdown": "Thanks!" }, { "id": 857333, "postDate": "2020-05-22T14:26:10.903Z", "content": "Do you know where can I find some kernel where someone uses BBoxSafeCrop? I cannot find it in the Albumentation's documentation.
", "rawMarkdown": "Do you know where can I find some kernel where someone uses BBoxSafeCrop? I cannot find it in the Albumentation's documentation. " }, { "id": 859878, "postDate": "2020-05-24T20:14:37.400Z", "content": "
Experiment：0.6860-->0.7111-->7161-->7191-->7286
", "rawMarkdown": "Experiment：0.6860-->0.7111-->7161-->7191-->7286", "replies": [ { "id": 853434, "postDate": "2020-05-19T07:22:57.887Z", "content": "Your baseline seems considerably high compared to mine. How many epochs do you train for? Is this FasterRCNN? Is this including pseudolabeling?
", "rawMarkdown": "Your baseline seems considerably high compared to mine. How many epochs do you train for? Is this FasterRCNN? Is this including pseudolabeling?" }, { "id": 855025, "postDate": "2020-05-20T14:13:54.907Z", "content": "I trained 22 epochs with faster-rcnn-r101,and no pseudolabeling.
", "rawMarkdown": "I trained 22 epochs with faster-rcnn-r101,and no pseudolabeling.", "votes": 1 }, { "id": 855305, "postDate": "2020-05-20T19:10:58.977Z", "content": "Hi @spytensor, this might be a silly question but I wanted to ask that are you using PyTorch faster-rccn with a different backbone or using some other implementation of FasterRCNN. Thanks
", "rawMarkdown": "Hi @spytensor, this might be a silly question but I wanted to ask that are you using PyTorch faster-rccn with a different backbone or using some other implementation of FasterRCNN. Thanks " }, { "id": 857070, "postDate": "2020-05-22T09:23:03.590Z", "content": "Hello! Do you use mmdetection to do this challenge? I'm new to DL, and I can't make a datareader to reader this data source in mmdet. Do you know how to use mmdet to read this data? Thanks!
", "rawMarkdown": "Hello! Do you use mmdetection to do this challenge? I'm new to DL, and I can't make a datareader to reader this data source in mmdet. Do you know how to use mmdet to read this data? Thanks!" } ] }, { "id": 862031, "postDate": "2020-05-26T11:29:12.170Z", "content": "Can you explain your opinion on why this works. I think its probably because of the difference in photos from different sources so this random brightness and saturation enables the model to learn the aspect of differing image quality. What do you think?
", "rawMarkdown": "Can you explain your opinion on why this works. I think its probably because of the difference in photos from different sources so this random brightness and saturation enables the model to learn the aspect of differing image quality. What do you think? ", "replies": [ { "id": 862100, "postDate": "2020-05-26T12:22:25.877Z", "content": "I think that's true but it is not only about image quality. You should consider the fact that photos are taken in different conditions (weather, sun, etc) and the wheat heads are quite different as well (color, etc). In plus augmentations, in general, are a good way to avoid overfitting.
", "rawMarkdown": "I think that's true but it is not only about image quality. You should consider the fact that photos are taken in different conditions (weather, sun, etc) and the wheat heads are quite different as well (color, etc). In plus augmentations, in general, are a good way to avoid overfitting. " } ] }, { "id": 857316, "postDate": "2020-05-22T14:15:05.667Z", "content": "Very Useful!
", "rawMarkdown": "Very Useful!" }, { "id": 854399, "postDate": "2020-05-20T02:24:17.830Z", "content": "what's your training size? As for me, 512 is better than 1024 on FasterRCNN.
", "rawMarkdown": "what's your training size? As for me, 512 is better than 1024 on FasterRCNN.", "replies": [ { "id": 855027, "postDate": "2020-05-20T14:14:43.253Z", "content": "I used 1024x1024.
", "rawMarkdown": "I used 1024x1024." } ] }, { "id": 854293, "postDate": "2020-05-20T00:04:08.187Z", "content": "Great job! I'm testing BrightnessContrast with HueSaturationValue, but my score decease, can these two trick work together? Or in one batch, I should use Oneof to choose one?
", "rawMarkdown": "Great job! I'm testing BrightnessContrast with HueSaturationValue, but my score decease, can these two trick work together? Or in one batch, I should use Oneof to choose one? ", "replies": [ { "id": 854372, "postDate": "2020-05-20T01:59:56.673Z", "content": "No, You should use these two methods separately.
", "rawMarkdown": "No, You should use these two methods separately.", "votes": 2 }, { "id": 854394, "postDate": "2020-05-20T02:17:51.403Z", "content": "Thanks!
", "rawMarkdown": "Thanks! " } ] }, { "id": 853550, "postDate": "2020-05-19T09:26:16.547Z", "content": "Have used image augmentation while working on image classifier. Found it to be useful mainly for minority classes.\nBy augmenting adding synthetic data model performance was improved.
", "rawMarkdown": "Have used image augmentation while working on image classifier. Found it to be useful mainly for minority classes.\nBy augmenting adding synthetic data model performance was improved. " }, { "id": 852300, "postDate": "2020-05-18T10:19:43.063Z", "content": "Does 1 really help? I would imagine that wheat heads are quite random in orientation in the images already.
", "rawMarkdown": "Does 1 really help? I would imagine that wheat heads are quite random in orientation in the images already.", "replies": [ { "id": 852352, "postDate": "2020-05-18T11:15:36.023Z", "content": "It does work.
", "rawMarkdown": "It does work." } ] }, { "id": 852226, "postDate": "2020-05-18T08:53:22.320Z", "content": "🙏 👍
", "rawMarkdown": "🙏 👍 " }, { "id": 852222, "postDate": "2020-05-18T08:44:35.303Z", "content": "What do you mean by multiscale test?
", "rawMarkdown": "What do you mean by multiscale test? ", "replies": [ { "id": 852232, "postDate": "2020-05-18T08:55:59.647Z", "content": "@matthewmasters use different scales to test single image,and merge them use nms.\nsuch as : [(512,512),(1024,1024)]
", "rawMarkdown": "@matthewmasters use different scales to test single image,and merge them use nms.\nsuch as : [(512,512),(1024,1024)]", "votes": 1 }, { "id": 857327, "postDate": "2020-05-22T14:23:23.703Z", "content": "@spytensor \nHi, is there any kernel/article etc. where I could read more about this technique?
", "rawMarkdown": "@spytensor \nHi, is there any kernel/article etc. where I could read more about this technique?", "votes": 1 }, { "id": 858296, "postDate": "2020-05-23T11:05:01.357Z", "content": "@spytensor Interesting. I actually found that ensembling two models in the fashion was worse than just my 1024^2 model.
", "rawMarkdown": "@spytensor Interesting. I actually found that ensembling two models in the fashion was worse than just my 1024^2 model." } ] }, { "id": 853136, "postDate": "2020-05-19T01:50:15.583Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 852626, "author_name": "Kaushal Shah", "author_url": "", "post_date": "2020-05-18T15:08:44.887000", "content": "@spytensor Thanks for sharing. I had one question: are you creating new larger dataset with this augmentation or just augmenting on the fly while training? I've tried some of these and my score decresed instead of increasing. I'm using original image size with 25 epochs of FasterRCNN.
", "votes": 1, "replies": [ { "id": 852971, "author_name": "Vivek Mittal", "author_url": "", "post_date": "2020-05-18T21:07:21.857000", "content": "I am also curious about this.🤔
", "votes": 0, "replies": [] }, { "id": 853142, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-19T01:57:37.517000", "content": "Randomly do augments while training models.
", "votes": 3, "replies": [] }, { "id": 853206, "author_name": "Kaushal Shah", "author_url": "", "post_date": "2020-05-19T03:06:46.323000", "content": "Cool! Thanks. @spytensor
", "votes": 0, "replies": [] } ] }, { "id": 852550, "author_name": "Oz", "author_url": "", "post_date": "2020-05-18T14:25:31.547000", "content": "Thanks for sharing! adding to your list, \nrandom sized BBoxSafeCrop worked for me, also using albumentations
", "votes": 1, "replies": [ { "id": 853144, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-19T01:59:19.773000", "content": "👍 ！Thanks ,I'll try it later.
", "votes": 0, "replies": [] }, { "id": 854290, "author_name": "27Wilson", "author_url": "", "post_date": "2020-05-20T00:01:45.697000", "content": "Hello! I'm testing RandomSizedBBoxSafeCrop, but sometimes it brings error which tells me the bbox number is 0. So hao to set he parameters of this trick especially erosion_rate?
", "votes": 0, "replies": [] }, { "id": 855371, "author_name": "Matt", "author_url": "", "post_date": "2020-05-20T21:37:02.343000", "content": "", "votes": 1, "replies": [] }, { "id": 857066, "author_name": "27Wilson", "author_url": "", "post_date": "2020-05-22T09:20:57.860000", "content": "Thanks!
", "votes": 0, "replies": [] }, { "id": 857333, "author_name": "Jan Malin", "author_url": "", "post_date": "2020-05-22T14:26:10.903000", "content": "Do you know where can I find some kernel where someone uses BBoxSafeCrop? I cannot find it in the Albumentation's documentation.
", "votes": 0, "replies": [] }, { "id": 859878, "author_name": "Fares FOURATI", "author_url": "", "post_date": "2020-05-24T20:14:37.400000", "content": "
Experiment：0.6860-->0.7111-->7161-->7191-->7286
", "votes": 0, "replies": [ { "id": 853434, "author_name": "Matt", "author_url": "", "post_date": "2020-05-19T07:22:57.887000", "content": "Your baseline seems considerably high compared to mine. How many epochs do you train for? Is this FasterRCNN? Is this including pseudolabeling?
", "votes": 0, "replies": [] }, { "id": 855025, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-20T14:13:54.907000", "content": "I trained 22 epochs with faster-rcnn-r101,and no pseudolabeling.
", "votes": 1, "replies": [] }, { "id": 855305, "author_name": "Vivek Mittal", "author_url": "", "post_date": "2020-05-20T19:10:58.977000", "content": "Hi @spytensor, this might be a silly question but I wanted to ask that are you using PyTorch faster-rccn with a different backbone or using some other implementation of FasterRCNN. Thanks
", "votes": 0, "replies": [] }, { "id": 857070, "author_name": "27Wilson", "author_url": "", "post_date": "2020-05-22T09:23:03.590000", "content": "Hello! Do you use mmdetection to do this challenge? I'm new to DL, and I can't make a datareader to reader this data source in mmdet. Do you know how to use mmdet to read this data? Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 862031, "author_name": "Amogh ", "author_url": "", "post_date": "2020-05-26T11:29:12.170000", "content": "Can you explain your opinion on why this works. I think its probably because of the difference in photos from different sources so this random brightness and saturation enables the model to learn the aspect of differing image quality. What do you think?
", "votes": 0, "replies": [ { "id": 862100, "author_name": "Fares FOURATI", "author_url": "", "post_date": "2020-05-26T12:22:25.877000", "content": "I think that's true but it is not only about image quality. You should consider the fact that photos are taken in different conditions (weather, sun, etc) and the wheat heads are quite different as well (color, etc). In plus augmentations, in general, are a good way to avoid overfitting.
", "votes": 0, "replies": [] } ] }, { "id": 857316, "author_name": "wittyenggs", "author_url": "", "post_date": "2020-05-22T14:15:05.667000", "content": "Very Useful!
", "votes": 0, "replies": [] }, { "id": 854399, "author_name": "Bet4Honor", "author_url": "", "post_date": "2020-05-20T02:24:17.830000", "content": "what's your training size? As for me, 512 is better than 1024 on FasterRCNN.
", "votes": 0, "replies": [ { "id": 855027, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-20T14:14:43.253000", "content": "I used 1024x1024.
", "votes": 0, "replies": [] } ] }, { "id": 854293, "author_name": "27Wilson", "author_url": "", "post_date": "2020-05-20T00:04:08.187000", "content": "Great job! I'm testing BrightnessContrast with HueSaturationValue, but my score decease, can these two trick work together? Or in one batch, I should use Oneof to choose one?
", "votes": 0, "replies": [ { "id": 854372, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-20T01:59:56.673000", "content": "No, You should use these two methods separately.
", "votes": 2, "replies": [] }, { "id": 854394, "author_name": "27Wilson", "author_url": "", "post_date": "2020-05-20T02:17:51.403000", "content": "Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 853550, "author_name": "Mansee", "author_url": "", "post_date": "2020-05-19T09:26:16.547000", "content": "Have used image augmentation while working on image classifier. Found it to be useful mainly for minority classes.\nBy augmenting adding synthetic data model performance was improved.
", "votes": 0, "replies": [] }, { "id": 852300, "author_name": "Cat Yip", "author_url": "", "post_date": "2020-05-18T10:19:43.063000", "content": "Does 1 really help? I would imagine that wheat heads are quite random in orientation in the images already.
", "votes": 0, "replies": [ { "id": 852352, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-18T11:15:36.023000", "content": "It does work.
", "votes": 0, "replies": [] } ] }, { "id": 852226, "author_name": "Tian", "author_url": "", "post_date": "2020-05-18T08:53:22.320000", "content": "🙏 👍
", "votes": 0, "replies": [] }, { "id": 852222, "author_name": "Matt", "author_url": "", "post_date": "2020-05-18T08:44:35.303000", "content": "What do you mean by multiscale test?
", "votes": 0, "replies": [ { "id": 852232, "author_name": "Spytensor", "author_url": "", "post_date": "2020-05-18T08:55:59.647000", "content": "@matthewmasters use different scales to test single image,and merge them use nms.\nsuch as : [(512,512),(1024,1024)]
", "votes": 1, "replies": [] }, { "id": 857327, "author_name": "Jan Malin", "author_url": "", "post_date": "2020-05-22T14:23:23.703000", "content": "@spytensor \nHi, is there any kernel/article etc. where I could read more about this technique?
", "votes": 1, "replies": [] }, { "id": 858296, "author_name": "Matt", "author_url": "", "post_date": "2020-05-23T11:05:01.357000", "content": "@spytensor Interesting. I actually found that ensembling two models in the fashion was worse than just my 1024^2 model.
", "votes": 0, "replies": [] } ] }, { "id": 853136, "author_name": "", "author_url": "", "post_date": "2020-05-19T01:50:15.583000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "852143": "1. random vertical-flip and horizontal-flip\n2. random BrightnessContrast\t\n3. random HueSaturationValue\n4. Multiscale Test\n\n1,2,3 use [albumentations](https://github.com/albumentations-team/albumentations)\n", "852626": "@spytensor Thanks for sharing. I had one question: are you creating new larger dataset with this augmentation or just augmenting on the fly while training? I've tried some of these and my score decresed instead of increasing. I'm using original image size with 25 epochs of FasterRCNN. ", "852550": "Thanks for sharing! adding to your list, \nrandom sized BBoxSafeCrop worked for me, also using albumentations ", "852234": "Experiment：0.6860-->0.7111-->7161-->7191-->7286", "862031": "Can you explain your opinion on why this works. I think its probably because of the difference in photos from different sources so this random brightness and saturation enables the model to learn the aspect of differing image quality. What do you think? ", "857316": "Very Useful!", "854399": "what's your training size? As for me, 512 is better than 1024 on FasterRCNN.", "854293": "Great job! I'm testing BrightnessContrast with HueSaturationValue, but my score decease, can these two trick work together? Or in one batch, I should use Oneof to choose one? ", "853550": "Have used image augmentation while working on image classifier. Found it to be useful mainly for minority classes.\nBy augmenting adding synthetic data model performance was improved. ", "852300": "Does 1 really help? I would imagine that wheat heads are quite random in orientation in the images already.", "852226": "🙏 👍 ", "852222": "What do you mean by multiscale test? ", "853136": "" } }