Kaggle: Carvana Image Masking Challenge
Vladimir Iglovikov
Data Scientist at Lyft
Kaggle Master
PhD in Physics
Kaggle: Carvana Image Masking Challenge
(1st out of 735)
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/E-1184x388.png)
Classification
![](http://html.scirp.org/file/4-7800353x6.png)
![](https://cs8.pikabu.ru/post_img/2016/11/09/6/1478684091112548169.jpg)
Cat (0.99)
Dog(0.01)
Classification to Segmentation: FCN8
![](https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/image_folder_7/FCN.jpg)
- Removed Dense Layers
- Added Simple Decoder (Upsampling + Conv)
FCN8 to SegNet
![](https://meetshah1995.github.io/images/blog/ss/fcn.png)
FCN8
SegNet
Replaced simple Decoder with Hierarchical Decoder
SegNet to UNet
![](http://cvlab.postech.ac.kr/research/deconvnet/images/overall.png)
SegNet
![](https://cdn-images-1.medium.com/max/953/1*Z98NhzbVISHa4CoemZS4Kw.png)
UNet
Added skip connections
Any binary Image Segmentation => UNet
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4101843/i164091.jpg)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4101848/i164091.png)
Medical Imaging
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4516119/JAX_Tile_017_RGB_0.jpg)
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4516120/JAX_Tile_017_RGB.png)
Satellite Imaging
MICCAI 2017 (Top medical imaging conference)
Robotic Instrument Segmentation
![](https://endovissub2017-roboticinstrumentsegmentation.grand-challenge.org/site/EndoVisSub2017-RoboticInstrumentSegmentation/serve/public_html/ims/banner.png)
9 research groups + our team.
4 evenings. 1st place.
(Paper in preparation)
Gastrointestinal Image ANAlysis (GIANA)
![](https://endovissub2017-giana.grand-challenge.org/site/EndoVisSub2017-GIANA/serve/public_html/segmentation_HD.png/)
11 research groups + our team
8 evenings. 1st place.
(2 papers in preparation)
Kaggle: Carvana Image Masking Challenge
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/E-1024x335.png)
- Train: 5088 Images
- Public test: 1200
- Private test: 3664
- Extra test: 95200
- Resolution 1918x1280
![](https://kaggle2.blob.core.windows.net/competitions/kaggle/3333/media/carvana_graphics.png)
Team
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/A-1024x483.png)
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/C-1024x482.png)
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/D-1024x488.png)
UNet modifications
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4516154/unet111.png)
UNet
UNet11
UNet with random weight initialization
UNet with pre-trained VGG11 encoder
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4516175/unet.png)
Metric and Loss Function
Binary cross entropy
![](https://upload.wikimedia.org/wikipedia/commons/thumb/1/1f/Intersection_of_sets_A_and_B.svg/371px-Intersection_of_sets_A_and_B.svg.png)
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/K-1024x342.png)
Augmentations and tricks
- flips
- scale
- rotation
- brightness
- contrast
- saturation
- conversion to grayscale
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/J-1024x677.png)
- Bagging
- Cyclic Learning Rate
Extra facts
![](http://5047-presscdn.pagely.netdna-cdn.com/wp-content/uploads/2017/12/B-768x776.png)
- Prize => $12,000
- Time spent ~10 evenings
- Time to train on one GPU => 90 days
- Time to predict => 13 days
- ~20 GPUs used
UNet is great but has issues
- Many classes => PSPNet, etc
- Slow and heavy => lighter encoders
- Does not catch sharp corners => better loss function
![](https://s3.amazonaws.com/media-p.slid.es/uploads/647006/images/4102382/mapillary.png)
66 classes
Blog post: Carvana Image Masking Challenge–1st Place Winner's Interview
Winning code: https://github.com/asanakoy/kaggle_carvana_segmentation
Weights for UNet11 trained on Carvana: https://github.com/ternaus/TernausNet
Title TextTitle Text
By Vladimir Iglovikov
Title TextTitle Text
- 3,641