Track 1: Single View Semantic 3D

An unrectified single-view image is provided for each geographic tile. The objective is to predict semantic labels and above-ground heights (meters).

Baseline Algorithm

For track 1, JHU/APL trained a U-Net for for image semantic segmentation and single-view depth prediction. These were run in Windows, but the code can just as easily run in Linux with minor edits to the params.py file which specifies folder locations.

Instructions for training and testing image semantic segmentation and single-view depth prediction models are provided in unets/README.md

Model weights

The model weights are now in a GitHub release zip file for download to avoid having large files in the code repo.

Quick Run

The following instructions are for running the baseline solution and assumes you have configured an Anaconda environment with dev-gpu.yml or dev-cpu.yml.

Data layout

Due do the necessary data packaging splits, to run the following algorithms you will need to point params.py's TRAIN_DIR to a folder containing the Track1 AGL, CLS, RGB and MSI tif files found in the Train-Track1-RGB, Train-Track1-MSI-*, and Train-Track1-Truth zip files. Similarly params.py's TEST_DIR should point to a folder containing MSI and RGB files found in the Validate-Track1 zip files.

Running

To train the UNET single view model, modify params.py with your file names and run:

python ./unets/runBaseline.py train single-view

To train train the UNET semantic model, modify params.py with your file names and run:

python ./unets/runBaseline.py train semantic

To test the semantic or single-view models, modify params.py with your file names and run:

python ./unets/runBaseline.py test semantic

or

python ./unets/runBaseline.py test single-view

Testing Solution

After running tests for the images in the validation set folder, the submission folder should have _CLS and _AGL files suitable for metric evaluation. Since you don't have the reference _CLS and _AGL files for the validation set used for leaderboard testing, you may try running the predictions above on hold-out images and then evaluate the leaderboard metrics on those files.

python .\metrics\track1.py [submission folder] [truth folder]

Here, the truth folder is a folder containing the holdouts, while the submission folder contains your predictions.