{"cells": [{"cell_type": "markdown", "source": ["# Summary table of the best solutions\n", "In this notebook I will write a summary of the posts with the best solutions to the challenge. "], "metadata": {"_uuid": "19caad72b069474d9df01fc375a09a16a237824d", "_cell_guid": "e94a8972-1937-4042-a0d0-a60e290702d6"}}, {"cell_type": "markdown", "source": ["| Position | name            | score  | number gpus | resolution                                  | models                                                                    | Cross-validation      | Data augmentation                                                                                                | Comments                                                                                                                                                                | link                                                                      |\n", "|----------|-----------------|--------|-------------|---------------------------------------------|---------------------------------------------------------------------------|-----------------------|------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------|\n", "| 3        | lyakaap         | 0.9972 | ?           | 1536x1024 & 1920x1280                       | U-Net + Dilated Conv                                                      | 5 folds \u00bfsplit?       | horizontal flip                                                                                                  | Pseudo Labeling                                                                                                                                                         | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40199 |\n", "| 5        | Kyle Y. Lee     | 0.9972 | 3           | multiple resolutions                        | Resnet-50 FCN                                                             | car split             | Horizontal flips, rotations (up to 10 degrees), height/width translation of about 5%, and zooms of about +/-10%. | Great effort on post-processing                                                                                                                                         | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40144 |\n", "| 6        | n01z3           | 0.9971 | 40          | full                                        | unet+LinkNet+pspnet                                                       | 5 folds \u00bfsplit?       |                                                                                                                  |                                                                                                                                                                         | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40119 |\n", "| 10       | David           | 0.9971 | 4           | full, splitted images in half for one model | unet                                                                      | ?                     |                                                                                                                  | Post processing on the antenas                                                                                                                                          | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40133 |\n", "| 11       | JandJ           | 0.9971 | ?           | Stage 1: 320x480, Stage 2: full             | unet                                                                      |                       | scale+-10%, flip left-right, brightness, contrast, pixel shift, etc                                              | Create a first prediction with 320x480 images, refine the prediction with 256x256 patches around the border                                                             | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40126 |\n", "| 15       | Andres Torrubia | 0.997  | ?           | full                                        | unet, unet with residual connections                                      | ?                     | random horizontal flips                                                                                          | first step to predict an initial mask with the full image, second step to refine the borders using patches, augmented input with background and the previous prediction | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40146 |\n", "| 23       | Aleksei Tiulpin | 0.997  | 4           | full, train on patches of 768x768           | Segnet based on VGG16 with batchnorm                                      | car split, 5 folds    | , rotation \u00b110 deg, horizontal flip, scaling \u00b120                                                                 |                                                                                                                                                                         | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40107 |\n", "| 30       | true_pk         | 0.9969 | 1           | 1024                                        | unet                                                                      | random                | https://www.kaggle.com/gaborfodor/augmentation-methods                                                           | First model to identify the car, crop the image and train 3 new models for better masks                                                                                 | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40123 |\n", "| 49       | fujisan         | 99.67  | 1           | 640x960                                     | customized Unet referred to Wide Residual Network, Squeeze-and-Excitation | one hold out 20% data | rotation, horizontal flip, scaling, HSV change, ganma correction, random erase                                   |                                                                                                                                                                         | https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40152 |"], "metadata": {"_uuid": "e064c199352a1057f7ddc4e4ae625090532e8545", "_cell_guid": "ac061283-70df-4b67-a861-f4f9d27eaa3e"}}, {"cell_type": "code", "source": [], "outputs": [], "execution_count": null, "metadata": {"_uuid": "97b5680d220b2543866d41726679304ca28ee0e7", "_cell_guid": "5c3d0be6-4bdb-4d32-a10d-67e594718a97", "collapsed": true}}], "nbformat": 4, "nbformat_minor": 1, "metadata": {"kernelspec": {"language": "python", "name": "python3", "display_name": "Python 3"}, "language_info": {"name": "python", "file_extension": ".py", "version": "3.6.1", "codemirror_mode": {"version": 3, "name": "ipython"}, "mimetype": "text/x-python", "pygments_lexer": "ipython3", "nbconvert_exporter": "python"}}}