What score you got for color images? I got 0.076. It's amazing that you got no better result when trained on color images compared with only trained on gray images. I guess you did some preprocessing on gray images when training or some on color images when testing?
", "rawMarkdown": "What score you got for color images? I got 0.076. It's amazing that you got no better result when trained on color images compared with only trained on gray images. I guess you did some preprocessing on gray images when training or some on color images when testing?", "replies": [ { "id": 304767, "postDate": "2018-03-28T01:50:59.023Z", "content": "Yes，I did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model.I got 0.070 for color image. How did you get 0.076 for color images? By training on color images only and testing? Did you use augmentation h&e?
", "rawMarkdown": "Yes，I did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model.I got 0.070 for color image. How did you get 0.076 for color images? By training on color images only and testing? Did you use augmentation h&e?" }, { "id": 304781, "postDate": "2018-03-28T02:19:26.397Z", "content": "I am using h&e, trained on all images. I guess post processing is important for my model
", "rawMarkdown": "I am using h&e, trained on all images. I guess post processing is important for my model" }, { "id": 304810, "postDate": "2018-03-28T03:22:57.280Z", "content": "Thanks, Yan. 0.076 is good result. How many validation data did you use for training model for color image?
", "rawMarkdown": "Thanks, Yan. 0.076 is good result. How many validation data did you use for training model for color image? " }, { "id": 307588, "postDate": "2018-04-02T04:49:39.670Z", "content": "Hi @xiapidan, you said
\n\n\n\n\nI did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model
\n
We have about 543 gray images, you trained these gray images to obtain a gray image model, then apply processing on color images (12 testing images) for using it in the testing of the gray model. Am I right? Which kind of processing on these color images did you use? I only got 0.06 on color image
", "rawMarkdown": "Hi @xiapidan, you said\n>I did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model\n\nWe have about 543 gray images, you trained these gray images to obtain a gray image model, then apply processing on color images (12 testing images) for using it in the testing of the gray model. Am I right? Which kind of processing on these color images did you use? I only got 0.06 on color image" }, { "id": 307646, "postDate": "2018-04-02T07:02:18.560Z", "content": "255 -（cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)）
", "rawMarkdown": "255 -（cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)）", "votes": 1 }, { "id": 308190, "postDate": "2018-04-03T05:10:59.743Z", "content": "@john1231983 @xiapidan Is your current score a score on the matterport model? Also, what kind of training schedule and aungmentation are you doing? \nI am using matterport code, but the score of the color image is 0.048.
", "rawMarkdown": "@john1231983 @xiapidan Is your current score a score on the matterport model? Also, what kind of training schedule and aungmentation are you doing? \nI am using matterport code, but the score of the color image is 0.048." }, { "id": 308785, "postDate": "2018-04-04T05:23:23.510Z", "content": "@ykamikawa: Yes, I used matterport with some post-processing like dilation to got 0.49. I trained on both gray and color and obtain 1 model and test these testing image on the model. When I measure the color socre only (extract from gray + color submission). I only got 0.055 LB.
\n\n@xiapidan: Thanks, I tried your suggestion but it does not improve my performance. Amazing with your score. I guess most of score are contributed by gray set. Do you apply some post-processing for gray set to obtain the good score in gray? To obtain 0.07LB on color, are you training on gray set (~510 images) and testing on color set?
", "rawMarkdown": "@ykamikawa: Yes, I used matterport with some post-processing like dilation to got 0.49. I trained on both gray and color and obtain 1 model and test these testing image on the model. When I measure the color socre only (extract from gray + color submission). I only got 0.055 LB. \n\n@xiapidan: Thanks, I tried your suggestion but it does not improve my performance. Amazing with your score. I guess most of score are contributed by gray set. Do you apply some post-processing for gray set to obtain the good score in gray? To obtain 0.07LB on color, are you training on gray set (~510 images) and testing on color set? " } ] }, { "id": 304280, "postDate": "2018-03-27T09:48:44.747Z", "content": "I trained color image only ,then test the color data in stage_test, I got no better result than training on gray image and test the color image \nI also tried to transfer the color image from BGR2HSV, then did training , got no better result either.
", "rawMarkdown": "I trained color image only ,then test the color data in stage_test, I got no better result than training on gray image and test the color image \nI also tried to transfer the color image from BGR2HSV, then did training , got no better result either.", "replies": [ { "id": 304285, "postDate": "2018-03-27T09:54:30.340Z", "content": "Which code based do you use? Heng code or matterport? I think image augmentation h&e may help
", "rawMarkdown": "Which code based do you use? Heng code or matterport? I think image augmentation h&e may help", "votes": 1 }, { "id": 304763, "postDate": "2018-03-28T01:43:18.623Z", "content": "Neither，and i also have tried to add h&e ，but no better result
", "rawMarkdown": "Neither，and i also have tried to add h&e ，but no better result" }, { "id": 304766, "postDate": "2018-03-28T01:50:08.630Z", "content": "Are you using Heng code or matterport? For matterport, you should add random crop in load_image_gt. And Have you try train with only 1 model, instead of 2? I found that no better result for separate training
", "rawMarkdown": "Are you using Heng code or matterport? For matterport, you should add random crop in load_image_gt. And Have you try train with only 1 model, instead of 2? I found that no better result for separate training", "votes": 2 }, { "id": 304776, "postDate": "2018-03-28T02:14:29.790Z", "content": "I am using matterport code. Adding random crop in load_image_gt works for you? I have tried training with only 1 model, no better result
", "rawMarkdown": "I am using matterport code. Adding random crop in load_image_gt works for you? I have tried training with only 1 model, no better result " }, { "id": 304795, "postDate": "2018-03-28T02:56:33.103Z", "content": "Yes. Random crop with size of 512. Otherwise you resize them then lost too much information. Random crop helps aug data also. I achieved 0.494 for matterport only using random crop , mosaic and h&e. How about you?
", "rawMarkdown": "Yes. Random crop with size of 512. Otherwise you resize them then lost too much information. Random crop helps aug data also. I achieved 0.494 for matterport only using random crop , mosaic and h&e. How about you?", "votes": 2 }, { "id": 305844, "postDate": "2018-03-29T13:59:21.920Z", "content": "Hi john. When you are using random crop, how did you set your IMAGE_MIN_DIM IMAGE_MAX_DIM in config file?
Just use 512. I was test 256 but it is very bad. I did not resize if the size is bigger than 512
", "rawMarkdown": "Just use 512. I was test 256 but it is very bad. I did not resize if the size is bigger than 512" }, { "id": 305936, "postDate": "2018-03-29T16:36:06.603Z", "content": "Hi John, how are you handling inference? Do you also take crops using a sliding window there or you just directly predict on a resized 512 image or full size? As I'm sure you noticed a lot of the test images are really narrow so resizing distorts the images a lot.
", "rawMarkdown": "Hi John, how are you handling inference? Do you also take crops using a sliding window there or you just directly predict on a resized 512 image or full size? As I'm sure you noticed a lot of the test images are really narrow so resizing distorts the images a lot. " }, { "id": 305945, "postDate": "2018-03-29T16:50:30.863Z", "content": "The second one, resize all to 512 and test on this size. I think it may have onther solution like sliding but it hard to merge cropped results to the final result
", "rawMarkdown": "The second one, resize all to 512 and test on this size. I think it may have onther solution like sliding but it hard to merge cropped results to the final result" }, { "id": 305986, "postDate": "2018-03-29T17:55:45.243Z", "content": "OK thanks! Yea I implemented a sliding window and restitch back the cropped results which seems to work as test results look better on window crop versus one-pass on the full resized image (especially on images with skewed aspect ratios). The only issue is like you say its difficult to keep track of merging the individual instance masks.
", "rawMarkdown": "OK thanks! Yea I implemented a sliding window and restitch back the cropped results which seems to work as test results look better on window crop versus one-pass on the full resized image (especially on images with skewed aspect ratios). The only issue is like you say its difficult to keep track of merging the individual instance masks. " }, { "id": 306113, "postDate": "2018-03-29T21:43:04.677Z", "content": "adding random_crop in load_image_gt did not improve the performance for me either.
", "rawMarkdown": "adding random_crop in load_image_gt did not improve the performance for me either." }, { "id": 306240, "postDate": "2018-03-30T05:14:44.013Z", "rawMarkdown": "", "isDeleted": true }, { "id": 306955, "postDate": "2018-03-31T12:07:29.303Z", "content": "It's amazing that get 0.49 only add random crop in matterport's maskrcnn!
\n\nI use matterport too, and I try several augmentation methods (Flip, Crop, GaussianBlur, rotates, sharpen) and add normalization for image loader
\n\nBut in addition to longer training time, there is no performance improvement
\n\nSo... Did you use some post/preprocess?
\n\nand what is “h&e”？Is that a augmentation method?
", "rawMarkdown": "It's amazing that get 0.49 only add random crop in matterport's maskrcnn!\n\nI use matterport too, and I try several augmentation methods (Flip, Crop, GaussianBlur, rotates, sharpen) and add normalization for image loader \n\nBut in addition to longer training time, there is no performance improvement\n\nSo... Did you use some post/preprocess?\n\nand what is “h&e”？Is that a augmentation method?" }, { "id": 306988, "postDate": "2018-03-31T13:54:47.123Z", "content": "Hi, John. When you use mosaic_train images to train your model, do you get the result csv file from mosaics_test?or just from stage1_test?
@john1231983, when you use random crop for matterport Mask RCNN, did you just do it during training and only for images larger than 512? In another word, did you still use resize when image are smaller than 512x512, and always use resize when detecting?
\n\nI feel resizing is hurting my performance pretty bad, but not sure how to exactly fix that. Can you give some hint? Thanks.
", "rawMarkdown": "@john1231983, when you use random crop for matterport Mask RCNN, did you just do it during training and only for images larger than 512? In another word, did you still use resize when image are smaller than 512x512, and always use resize when detecting? \n\nI feel resizing is hurting my performance pretty bad, but not sure how to exactly fix that. Can you give some hint? Thanks.\n" }, { "id": 307516, "postDate": "2018-04-01T22:37:31.610Z", "content": "i did not test on mosaic testing set. I only train on mosaic training set and test on original image. First, i will random crop 512x512 if the size of image is bigger than 512 ( not using prob crop), otherwise resize image to 512x512. I trained with 60 epochs on heads and 40 epoch on alls with learning rate 0.0001 using Adam. I don't know why someone success to train with SGD( i used sgd but got 0.44). Using above suggestion, you may got 0.47 (no post)~0.49lb (with post processing)
", "rawMarkdown": "i did not test on mosaic testing set. I only train on mosaic training set and test on original image. First, i will random crop 512x512 if the size of image is bigger than 512 ( not using prob crop), otherwise resize image to 512x512. I trained with 60 epochs on heads and 40 epoch on alls with learning rate 0.0001 using Adam. I don't know why someone success to train with SGD( i used sgd but got 0.44). Using above suggestion, you may got 0.47 (no post)~0.49lb (with post processing)", "votes": 1 }, { "id": 307517, "postDate": "2018-04-01T22:49:14.550Z", "content": "Thank you for your quick and detailed response! I will try it out. One extra question, when you do need to resize, did you use the padding method (as in matterport's resize_image, which probably preserves the height/width ratio), or did you use the skimage.transform.resize?
", "rawMarkdown": "Thank you for your quick and detailed response! I will try it out. One extra question, when you do need to resize, did you use the padding method (as in matterport's resize_image, which probably preserves the height/width ratio), or did you use the skimage.transform.resize? " } ] } ], "comments": [ { "id": 304382, "author_name": "Yan Wang", "author_url": "", "post_date": "2018-03-27T13:45:40.600000", "content": "What score you got for color images? I got 0.076. It's amazing that you got no better result when trained on color images compared with only trained on gray images. I guess you did some preprocessing on gray images when training or some on color images when testing?
", "votes": 0, "replies": [ { "id": 304767, "author_name": "xiapidan", "author_url": "", "post_date": "2018-03-28T01:50:59.023000", "content": "Yes，I did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model.I got 0.070 for color image. How did you get 0.076 for color images? By training on color images only and testing? Did you use augmentation h&e?
", "votes": 0, "replies": [] }, { "id": 304781, "author_name": "Yan Wang", "author_url": "", "post_date": "2018-03-28T02:19:26.397000", "content": "I am using h&e, trained on all images. I guess post processing is important for my model
", "votes": 0, "replies": [] }, { "id": 304810, "author_name": "zmingen", "author_url": "", "post_date": "2018-03-28T03:22:57.280000", "content": "Thanks, Yan. 0.076 is good result. How many validation data did you use for training model for color image?
", "votes": 0, "replies": [] }, { "id": 307588, "author_name": "john1231983", "author_url": "", "post_date": "2018-04-02T04:49:39.670000", "content": "Hi @xiapidan, you said
\n\n\n\n\nI did some processing on color images when testing by gray image model and got better result than with no processing on color images when testing by color image model
\n
We have about 543 gray images, you trained these gray images to obtain a gray image model, then apply processing on color images (12 testing images) for using it in the testing of the gray model. Am I right? Which kind of processing on these color images did you use? I only got 0.06 on color image
", "votes": 0, "replies": [] }, { "id": 307646, "author_name": "xiapidan", "author_url": "", "post_date": "2018-04-02T07:02:18.560000", "content": "255 -（cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)）
", "votes": 1, "replies": [] }, { "id": 308190, "author_name": "ykamikawa", "author_url": "", "post_date": "2018-04-03T05:10:59.743000", "content": "@john1231983 @xiapidan Is your current score a score on the matterport model? Also, what kind of training schedule and aungmentation are you doing? \nI am using matterport code, but the score of the color image is 0.048.
", "votes": 0, "replies": [] }, { "id": 308785, "author_name": "john1231983", "author_url": "", "post_date": "2018-04-04T05:23:23.510000", "content": "@ykamikawa: Yes, I used matterport with some post-processing like dilation to got 0.49. I trained on both gray and color and obtain 1 model and test these testing image on the model. When I measure the color socre only (extract from gray + color submission). I only got 0.055 LB.
\n\n@xiapidan: Thanks, I tried your suggestion but it does not improve my performance. Amazing with your score. I guess most of score are contributed by gray set. Do you apply some post-processing for gray set to obtain the good score in gray? To obtain 0.07LB on color, are you training on gray set (~510 images) and testing on color set?
", "votes": 0, "replies": [] } ] }, { "id": 304280, "author_name": "xiapidan", "author_url": "", "post_date": "2018-03-27T09:48:44.747000", "content": "I trained color image only ,then test the color data in stage_test, I got no better result than training on gray image and test the color image \nI also tried to transfer the color image from BGR2HSV, then did training , got no better result either.
", "votes": 0, "replies": [ { "id": 304285, "author_name": "john1231983", "author_url": "", "post_date": "2018-03-27T09:54:30.340000", "content": "Which code based do you use? Heng code or matterport? I think image augmentation h&e may help
", "votes": 1, "replies": [] }, { "id": 304763, "author_name": "xiapidan", "author_url": "", "post_date": "2018-03-28T01:43:18.623000", "content": "Neither，and i also have tried to add h&e ，but no better result
", "votes": 0, "replies": [] }, { "id": 304766, "author_name": "john1231983", "author_url": "", "post_date": "2018-03-28T01:50:08.630000", "content": "Are you using Heng code or matterport? For matterport, you should add random crop in load_image_gt. And Have you try train with only 1 model, instead of 2? I found that no better result for separate training
", "votes": 2, "replies": [] }, { "id": 304776, "author_name": "xiapidan", "author_url": "", "post_date": "2018-03-28T02:14:29.790000", "content": "I am using matterport code. Adding random crop in load_image_gt works for you? I have tried training with only 1 model, no better result
", "votes": 0, "replies": [] }, { "id": 304795, "author_name": "john1231983", "author_url": "", "post_date": "2018-03-28T02:56:33.103000", "content": "Yes. Random crop with size of 512. Otherwise you resize them then lost too much information. Random crop helps aug data also. I achieved 0.494 for matterport only using random crop , mosaic and h&e. How about you?
", "votes": 2, "replies": [] }, { "id": 305844, "author_name": "Zheng Li", "author_url": "", "post_date": "2018-03-29T13:59:21.920000", "content": "Hi john. When you are using random crop, how did you set your IMAGE_MIN_DIM IMAGE_MAX_DIM in config file?
Just use 512. I was test 256 but it is very bad. I did not resize if the size is bigger than 512
", "votes": 0, "replies": [] }, { "id": 305936, "author_name": "James Requa", "author_url": "", "post_date": "2018-03-29T16:36:06.603000", "content": "Hi John, how are you handling inference? Do you also take crops using a sliding window there or you just directly predict on a resized 512 image or full size? As I'm sure you noticed a lot of the test images are really narrow so resizing distorts the images a lot.
", "votes": 0, "replies": [] }, { "id": 305945, "author_name": "john1231983", "author_url": "", "post_date": "2018-03-29T16:50:30.863000", "content": "The second one, resize all to 512 and test on this size. I think it may have onther solution like sliding but it hard to merge cropped results to the final result
", "votes": 0, "replies": [] }, { "id": 305986, "author_name": "James Requa", "author_url": "", "post_date": "2018-03-29T17:55:45.243000", "content": "OK thanks! Yea I implemented a sliding window and restitch back the cropped results which seems to work as test results look better on window crop versus one-pass on the full resized image (especially on images with skewed aspect ratios). The only issue is like you say its difficult to keep track of merging the individual instance masks.
", "votes": 0, "replies": [] }, { "id": 306113, "author_name": "Mehul Sampat", "author_url": "", "post_date": "2018-03-29T21:43:04.677000", "content": "adding random_crop in load_image_gt did not improve the performance for me either.
", "votes": 0, "replies": [] }, { "id": 306240, "author_name": "", "author_url": "", "post_date": "2018-03-30T05:14:44.013000", "content": "", "votes": 0, "replies": [] }, { "id": 306955, "author_name": "Caibimiao", "author_url": "", "post_date": "2018-03-31T12:07:29.303000", "content": "It's amazing that get 0.49 only add random crop in matterport's maskrcnn!
\n\nI use matterport too, and I try several augmentation methods (Flip, Crop, GaussianBlur, rotates, sharpen) and add normalization for image loader
\n\nBut in addition to longer training time, there is no performance improvement
\n\nSo... Did you use some post/preprocess?
\n\nand what is “h&e”？Is that a augmentation method?
", "votes": 0, "replies": [] }, { "id": 306988, "author_name": "Zheng Li", "author_url": "", "post_date": "2018-03-31T13:54:47.123000", "content": "Hi, John. When you use mosaic_train images to train your model, do you get the result csv file from mosaics_test?or just from stage1_test?
@john1231983, when you use random crop for matterport Mask RCNN, did you just do it during training and only for images larger than 512? In another word, did you still use resize when image are smaller than 512x512, and always use resize when detecting?
\n\nI feel resizing is hurting my performance pretty bad, but not sure how to exactly fix that. Can you give some hint? Thanks.
", "votes": 0, "replies": [] }, { "id": 307516, "author_name": "john1231983", "author_url": "", "post_date": "2018-04-01T22:37:31.610000", "content": "i did not test on mosaic testing set. I only train on mosaic training set and test on original image. First, i will random crop 512x512 if the size of image is bigger than 512 ( not using prob crop), otherwise resize image to 512x512. I trained with 60 epochs on heads and 40 epoch on alls with learning rate 0.0001 using Adam. I don't know why someone success to train with SGD( i used sgd but got 0.44). Using above suggestion, you may got 0.47 (no post)~0.49lb (with post processing)
", "votes": 1, "replies": [] }, { "id": 307517, "author_name": "Rui Li", "author_url": "", "post_date": "2018-04-01T22:49:14.550000", "content": "Thank you for your quick and detailed response! I will try it out. One extra question, when you do need to resize, did you use the padding method (as in matterport's resize_image, which probably preserves the height/width ratio), or did you use the skimage.transform.resize?
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "303329": "", "304382": "What score you got for color images? I got 0.076. It's amazing that you got no better result when trained on color images compared with only trained on gray images. I guess you did some preprocessing on gray images when training or some on color images when testing?", "304280": "I trained color image only ,then test the color data in stage_test, I got no better result than training on gray image and test the color image \nI also tried to transfer the color image from BGR2HSV, then did training , got no better result either." } }