Skip to content

Latest commit

 

History

History
122 lines (72 loc) · 6.98 KB

EXPERIMENTS.md

File metadata and controls

122 lines (72 loc) · 6.98 KB
Raw

Experiments

Here are some of the training datasets and trained models that have come out of our experiments. Trained weights (.caffemodel files) are available in the requester-pays S3 bucket skynet-models -- see below for specific paths.

Training Data

The license for the satellite/aerial imagery we used means that we cannot redistribute the files ourselves, but the skynet-data scripts should hopefully make it easier for you to create your own. (If you want to reproduce one of these exactly, see the /data directory in this repo for a list of specific map tiles used for each training set.)

Set N Resolution Labels Source Location
training-set-7 3954 z16 road (stroke, 5px) Mapbox Satellite Northeastern USA and some of CAN
training-set-8 24747 z17 water (fill)
building (fill)
road (stroke, 5px)		 Mapbox Satellite Continental USA
training-set-9 6747 z17 road (stroke, 5px) Mapbox Satellite Seattle, USA
training-set-10 6914 z17 water (fill)
building (fill)
road (stroke, 5px)		 Mapbox Satellite Seattle, USA
training-set-11 6841 z17 building (fill)
road (stroke, 5px)		 Mapbox Satellite Seattle, WA, USA
training-set-12 6450 z17 building (fill)
road (stroke, 5px)		 Mapbox Satellite Dar es Salaam, TZA
spacenet-1 3075 50cm building (fill) SpaceNet on AWS Rio De Janeiro, BRA

Trained Models

A model with 'tsX' in the name corresponds to to 'training set X' above.

segnet-ts7-1

Trained caffemodel files: s3://skynet-models/segnet-ts7-1/manifest.txt

Iterations Correctness Completeness View Results
5000 0.3256573519036912 0.7868625456307186 view results

segnet-ts7-2

Trained caffemodel files: s3://skynet-models/segnet-ts7-2/manifest.txt

Iterations Correctness Completeness View Results
60000 0.5058499453227574 0.6716329348245725 view results

segnet-ts9-1

Trained caffemodel files: s3://skynet-models/segnet-ts9-1/manifest.txt

Compare results

Iterations Correctness Completeness View Results
5000 0.518253184652922 0.8573735411419057 view results
55000 0.7707086613419863 0.7196262477292044 view results

segnet-ts10-1

Trained caffemodel files: s3://skynet-models/segnet-ts10-1/manifest.txt

Compare results

Iterations Correctness Completeness View Results
20K 0.427194953622606 0.46094955169178686 view results
40K 0.5728470881115866 0.638696209213571 view results

segnet-ts11-1

Trained caffemodel files: s3://skynet-models/segnet-ts11-1/manifest.txt

Compare results

Iterations Correctness Completeness View Results
10K 0.6246305164870534 0.6131471814748839 view results
30K 0.7283672440907986 0.45904370319881704 view results

segnet-ts12-1

Trained caffemodel files: s3://skynet-models/segnet-ts12-1/manifest.txt

Compare results

Iterations Correctness Completeness View Results
15K 0.6706866344254577 0.8394262122454237 view results
25K 0.6006667123465258 0.8542909458638581 view results
40K 0.6372450268591862 0.8168847046209704 view results

spacenet-1.0

Trained caffemodel files: s3://skynet-models/spacenet-1.0/manifest.txt

Compare results

Iterations Correctness Completeness View Results
5000 0.7884645371187057 0.9409724819069966 view results
10000 0.7882820153922176 0.9430313923848284 view results
15000 0.7902352382903021 0.9457690149692252 view results
20000 0.7918245263758488 0.9426878128090946 view results
25000 0.7923963798908541 0.9405211082529236 view results