CMPUT414

CMPUT 414 course (Winter 2019) Team Hawaii - Leland Jansen and Nathan Liebrecht

Reproducing results

Our project is grouped into 3 broad steps:

Generate the 64x64 raster tiles
Train the YOLO CNN
Testing inference

To get pre-generated training data (data.zip) and final trained weights download the files at: https://drive.google.com/drive/u/1/folders/10x4pRpnZcyZsu-ysS38-0K1WWxSCFhEj

Pre-requisits

The following python libraries must be installed:

sudo pip3 -U install pptk numpy sklearn

Next, Darknet must be downloaded and compiled. Simply clone the darknet repository into the src folder and compile it. The compilation instructions and code repository can be found here: https://github.com/AlexeyAB/darknet#how-to-compile-on-linux

LIBSO=1 must be enabled for inference to work.

1. Generating the 64x64 raster tiles

To skip this step, simply extract the data.zip folder into the src/darknet/build/darknet/x64 folder created in the "Pre-requisits" section.

To re-rasterize the Sydney dataset included in the data folder into the 64x64 tiles

Download the Sydney Urban Objects Dataset and extract the data into the data/sydney-urban-objects-dataset/ folder. The dataset can be found here: http://www.acfr.usyd.edu.au/papers/SydneyUrbanObjectsDataset.shtml
Create an out folder in the root of the repository
Run visualize.py. Data will be output in the out folder. This will take around 4 hours. Once this is complete, copy the files in out into a new folder: src/darknet/build/darknet/x64/data/414
Run split_dataset.py to regenerate the data split. Copy test.txt and train.txt into the src/darknet/build/darknet/x64/data folder.
Copy the remaining configuration files into the data folder from the data.zip file. These are fixed configuration files that remain constant

You should now have a file structure similar to the provided data.zip file.

2. Train the YOLO CNN

To skip this step you may just copy over the yolov3-414_final.weights found in the drive link into src/darknet/build/darknet/x64.

To train from scratch:

Change to the directory with the darknet binary (src/darknet/build/darknet/x64)
Download the starting weights https://pjreddie.com/media/files/darknet53.conv.74
Run darknet detector train data/414.data data/yolov3-414.cfg darknet53.conv.74 -map
Move src/darknet/build/darknet/x64/backup/yolov3-414_final.weights up a directory.

3. Testing inference