KDD-BR 2022

PaRaNA solution for KDD-BR 2022 (Visual odometry for autonomous UAVs in GNSS contested environments)

The steps to reproduce our solution are as follows:

1. Manually download and extract the competition .zip from Kaggle into the kddbr-2022 folder of this repository. You must also manually extract the prediction zips found in predictions/.

2. Generate Features using PDCNet model. These are already precomputed and found at features/PDCNet. If you actually want to generate them yourself from scratch, use the notebook 0) Generate features from scratch using DenseMatching library.ipynb. Note that this requires you to download the DenseMatching repository, and follow its instructions to set up the correct environment before running the .ipynb file.

3. Create the 'flight' CSVs. These CSVs are our attempt to restore temporal information that was lost in the original input data. They are already precomputed, so you can just use the ones at found at flight/. Alternatively, they may be generated from scratch using notebook 1) Infer Flight order, create Flight csvs.ipynb. The trick is to:

   • Start at some random row of public.csv DataFrame.
   • Calculate new_altitude = altitude + delta
   • Check for exact matches for the new_altitude. If 0 or more than 1 matches are found, then we filter the 100 closest altitudes and use mean squared error between images to determine the closest match.
   • Repeat process until all rows are visited.

4. Use LightGBM to map the Extracted Features to the output values (separate Boosters for North, East variables). If you want to train it from scratch, use 2) Gradient Boosting using DenseMatching Features.ipynb . To run this script, it should be sufficient to run pip install lightgbm==3.3.2 pandas matplotlib tqdm. Do note that, because of random parameters such as feature fraction, you might end up with a slightly different model. To take that into account, we have also provided the files for the actual Booster model used in the final prediction. Notebook 3) Verification - Gradient Boosting using DenseMatching Features-Oct23.ipynb verifies that the predictions of this saved model on the test set does indeed match exactly the one that was sent to Kaggle.