Hi all,\nI wanted to ask for some help on this RLE encoding. After running the test data through my trained network and viewing the input and predicted masks side by side the results look VERY accurate along with a very good IOU and loss on the val data.(obviously the eye doesnt have the ability to compute iou, so i may just be wrong). After using the encoding technique in one of the highest ranked Kernals my output csv look like the attached. This scored a .033 so I was curious if the the ids may be in incorrect order or if the RLE process was wrong. Thanks in advance!
", "messages": [ { "id": "277377", "postDate": "02/02/2018 23:22:01", "content": "Hi all,\nI wanted to ask for some help on this RLE encoding. After running the test data through my trained network and viewing the input and predicted masks side by side the results look VERY accurate along with a very good IOU and loss on the val data.(obviously the eye doesnt have the ability to compute iou, so i may just be wrong). After using the encoding technique in one of the highest ranked Kernals my output csv look like the attached. This scored a .033 so I was curious if the the ids may be in incorrect order or if the RLE process was wrong. Thanks in advance!
", "rawMarkdown": "Hi all,\nI wanted to ask for some help on this RLE encoding. After running the test data through my trained network and viewing the input and predicted masks side by side the results look VERY accurate along with a very good IOU and loss on the val data.(obviously the eye doesnt have the ability to compute iou, so i may just be wrong). After using the encoding technique in one of the highest ranked Kernals my output csv look like the attached. This scored a .033 so I was curious if the the ids may be in incorrect order or if the RLE process was wrong. Thanks in advance!", "votes": null }, { "id": "277407", "postDate": "02/03/2018 02:05:39", "content": "There is no problem with the evaluation. From quickly going through your csv, It can be clearly seen that it is quite low-scoring. Either the predicted images are very noisy (i.e many values are only 1 pixel '21590 1') or you just pass wrong predictions to the encoding code which generated RLE.
", "rawMarkdown": "There is no problem with the evaluation. From quickly going through your csv, It can be clearly seen that it is quite low-scoring. Either the predicted images are very noisy (i.e many values are only 1 pixel '21590 1') or you just pass wrong predictions to the encoding code which generated RLE.", "votes": null }, { "id": "277866", "postDate": "02/04/2018 16:46:04", "content": "Thank You @Oleksandr. I'm starting to become skeptical of passing the correct images to the encoding. I have tried multiple network architectures and have applied a cleaning function to all of the images and seem to get a .033 IOU every time. Are you or anyone else familiar with a technique to decode the csv into images in order to view if the images are correctly encoded
", "rawMarkdown": "Thank You @Oleksandr. I'm starting to become skeptical of passing the correct images to the encoding. I have tried multiple network architectures and have applied a cleaning function to all of the images and seem to get a .033 IOU every time. Are you or anyone else familiar with a technique to decode the csv into images in order to view if the images are correctly encoded", "votes": null }, { "id": "277914", "postDate": "02/04/2018 20:04:24", "content": "i just posted a discussion in it. i think its connected to that that the data masks was dots seperated. so any segmentation wont get high score because some dots are connected to each other and u need to seperate them in the rle encoding. so i think or u find a way to encode them seperate or u need to go with the detection methods bbox everybody talking about
", "rawMarkdown": "i just posted a discussion in it. i think its connected to that that the data masks was dots seperated. so any segmentation wont get high score because some dots are connected to each other and u need to seperate them in the rle encoding. so i think or u find a way to encode them seperate or u need to go with the detection methods bbox everybody talking about", "votes": null } ], "comments": [ { "id": 277407, "author_name": "ukrdailo", "author_url": "", "post_date": "02/03/2018 02:05:39", "content": "There is no problem with the evaluation. From quickly going through your csv, It can be clearly seen that it is quite low-scoring. Either the predicted images are very noisy (i.e many values are only 1 pixel '21590 1') or you just pass wrong predictions to the encoding code which generated RLE.
", "votes": null, "replies": [] }, { "id": 277866, "author_name": "ianrowan", "author_url": "", "post_date": "02/04/2018 16:46:04", "content": "Thank You @Oleksandr. I'm starting to become skeptical of passing the correct images to the encoding. I have tried multiple network architectures and have applied a cleaning function to all of the images and seem to get a .033 IOU every time. Are you or anyone else familiar with a technique to decode the csv into images in order to view if the images are correctly encoded
", "votes": null, "replies": [ { "id": 277914, "author_name": "aviwolfson", "author_url": "", "post_date": "02/04/2018 20:04:24", "content": "i just posted a discussion in it. i think its connected to that that the data masks was dots seperated. so any segmentation wont get high score because some dots are connected to each other and u need to seperate them in the rle encoding. so i think or u find a way to encode them seperate or u need to go with the detection methods bbox everybody talking about
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "277377": "Hi all,\nI wanted to ask for some help on this RLE encoding. After running the test data through my trained network and viewing the input and predicted masks side by side the results look VERY accurate along with a very good IOU and loss on the val data.(obviously the eye doesnt have the ability to compute iou, so i may just be wrong). After using the encoding technique in one of the highest ranked Kernals my output csv look like the attached. This scored a .033 so I was curious if the the ids may be in incorrect order or if the RLE process was wrong. Thanks in advance!", "277407": "There is no problem with the evaluation. From quickly going through your csv, It can be clearly seen that it is quite low-scoring. Either the predicted images are very noisy (i.e many values are only 1 pixel '21590 1') or you just pass wrong predictions to the encoding code which generated RLE.", "277866": "Thank You @Oleksandr. I'm starting to become skeptical of passing the correct images to the encoding. I have tried multiple network architectures and have applied a cleaning function to all of the images and seem to get a .033 IOU every time. Are you or anyone else familiar with a technique to decode the csv into images in order to view if the images are correctly encoded", "277914": "i just posted a discussion in it. i think its connected to that that the data masks was dots seperated. so any segmentation wont get high score because some dots are connected to each other and u need to seperate them in the rle encoding. so i think or u find a way to encode them seperate or u need to go with the detection methods bbox everybody talking about" }, "source": "meta" }