• Using the IceVision library for object detection, I built an object detector to detect buildings from the AICrowd Mapping Challenge dataset.

• Ultralytics YOLOv5, pretrained on the COCO dataset, obtained a COCO metric of 57%, despite this dataset being dissimilar to COCO.

• The model recalled 76% of the bounding boxes with a precision of 85%.

See below for an inference example.

The model training notebook in this repository can be run in Google Colab or Paperspace, while the inference notebook using Voxel Fifty One should be run locally. The instructions shall be described below.

First download and decompress the Mapping Challenge data from here into the "data" folder.

The notebooks in this repo can be run after installing IceVision in a bash environment:

# # Torch - Torchvision - IceVision - IceData - MMDetection - YOLOv5 - EfficientDet Installation
wget https://raw.githubusercontent.com/airctic/icevision/master/icevision_install.sh

# # Choose your installation target: cuda11 or cuda10 or cpu
source icevision_install.sh cuda11 master

They've been tested on July 2022.

Once the dataset is downloaded, and IceVision is installed.

1. Preprocess annotations-small.json so that it works well with IceVision by running the following line:

python mappingchallenge/preprocess-annotations.py data/train annotation-small.json annotation-small3.json

For testing purposes we will use the validation set, so replace "data/train" with "data/val" above.

2. Create a cache of the IceVision records, since parsing 8000 images takes some time:

python mappingchallenge/cache-records.py data/train annotation-small3.json cache-records3.p

For testing purposes we will use the validation set, so replace "data/train" with "data/val" above.

3. Configure and train models in "notebooks/wandb-building-detection.ipynb". Ideally you'll need a Weights & Biases (wandb) account to get this working.

4. Analyze model performance using the notebook "notebooks/inference-building-detection.ipynb". Note that this notebook should be run locally, and is not guarenteed to work on Colab or Paperspace.

This can be done without training a model by unzipping the model in the models folder.