DSTL

https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/

Object types

Note that labels here are 1 less than in submission file:

• 0: Buildings - large building, residential, non-residential, fuel storage facility, fortified building
• 1: Misc. Manmade structures
• 2: Road
• 3: Track - poor/dirt/cart track, footpath/trail
• 4: Trees - woodland, hedgerows, groups of trees, standalone trees
• 5: Crops - contour ploughing/cropland, grain (wheat) crops, row (potatoes, turnips) crops
• 6: Waterway
• 7: Standing water
• 8: Vehicle Large - large vehicle (e.g. lorry, truck,bus), logistics vehicle
• 9: Vehicle Small - small vehicle (car, van), motorbike

Object types stats

Area:

-------- -------- ------- ------- -------- -------- -------- -------- ------- ------- ------- im_id 0 1 2 3 4 5 6 7 8 9 -------- -------- ------- ------- -------- -------- -------- -------- ------- ------- ------- 6010_1_2 0.0% 0.0653% 0.0% 1.3345% 4.5634% 0.0% 0.0% 0.0% 0.0% 0.0% 6010_4_2 0.0% 0.0% 0.0% 1.9498% 12.3410% 0.0% 0.0% 0.0% 0.0% 0.0% 6010_4_4 0.0% 0.0% 0.0% 0.0% 22.8556% 0.0% 0.0% 0.0% 0.0% 0.0% 6040_1_0 0.0% 0.0% 0.0% 1.4446% 8.0062% 0.0% 0.0% 0.0% 0.0% 0.0% 6040_1_3 0.0% 0.0% 0.0% 0.2019% 18.7376% 3.6610% 0.0% 0.0% 0.0% 0.0% 6040_2_2 0.0% 0.0% 0.0% 0.9581% 18.7348% 0.0% 0.0% 0.0% 0.0% 0.0% 6040_4_4 0.0% 0.0% 0.0% 1.8893% 2.9152% 0.0% 0.0% 0.0% 0.0% 0.0% 6060_2_3 0.1389% 0.3037% 0.0% 3.0302% 8.4519% 93.5617% 0.0% 0.0% 0.0% 0.0003% 6070_2_3 1.5524% 0.3077% 0.8135% 0.0% 16.0439% 0.0% 10.6325% 0.0543% 0.0% 0.0058% 6090_2_0 0.0% 0.0343% 0.0% 0.4072% 10.1105% 28.2399% 0.0% 0.3130% 0.0% 0.0008% 6100_1_3 8.7666% 2.7289% 2.2145% 12.2506% 6.2015% 2.6901% 0.0% 0.6839% 0.0110% 0.0459% 6100_2_2 3.1801% 0.8188% 1.1903% 3.7222% 7.6089% 44.3148% 1.8823% 0.0512% 0.0100% 0.0242% 6100_2_3 8.2184% 1.4110% 1.2099% 9.5948% 7.5323% 0.0% 0.0% 0.0603% 0.0148% 0.0661% 6110_1_2 13.1314% 2.8616% 0.4192% 4.1817% 3.3154% 49.7792% 0.0% 0.1527% 0.0% 0.0065% 6110_3_1 4.5495% 1.2561% 3.6302% 2.8221% 5.4133% 57.6089% 0.0% 0.5531% 0.0181% 0.0253% 6110_4_0 2.4051% 0.5732% 1.8409% 2.8067% 5.7379% 80.7666% 0.0% 1.4210% 0.0136% 0.0017% 6120_2_0 1.7980% 0.7257% 0.8505% 4.4026% 5.6352% 79.5910% 0.0% 0.0% 0.0138% 0.0041% 6120_2_2 20.6570% 2.0389% 4.2547% 8.6533% 4.4347% 10.2929% 0.0% 0.2859% 0.0076% 0.1560% 6140_1_2 12.9211% 2.4488% 0.3538% 4.1461% 3.1027% 49.5910% 0.0% 0.1415% 0.0% 0.0086% 6140_3_1 5.2015% 1.4349% 3.4252% 2.5189% 5.8852% 57.3959% 0.0% 0.4664% 0.0042% 0.0358% 6150_2_3 0.0% 0.6055% 0.0% 3.0197% 13.5187% 80.6649% 0.0% 0.0% 0.0% 0.0% 6160_2_1 0.0% 0.0% 0.0% 2.7986% 10.2713% 0.0% 0.0% 0.0% 0.0% 0.0% 6170_0_4 0.0% 0.0016% 0.0% 0.1994% 24.8913% 0.0% 0.0% 0.0152% 0.0% 0.0% 6170_2_4 0.0% 0.0011% 0.0% 2.5070% 7.7844% 49.5326% 0.0% 0.0089% 0.0% 0.0% 6170_4_1 0.0% 0.0% 0.0% 0.1349% 20.2214% 0.0% 0.0% 0.0% 0.0% 0.0% -------- -------- ------- ------- -------- -------- -------- -------- ------- ------- -------

Making submission

Train a CNN (choose number of epochs and other hyper-params running without --all):

$ ./train.py debug12-all-ep4 --all --hps n_epochs=4

TODO: document train_all, choosing epsilon with make_submission --train-only

Make submission file (be sure to pass hyperparameters that influence CNN structure):

$ ./make_submission.py debug12-all-ep4 sub_th0.4_debug12-all-ep4 --threshold 0.4