Quick attempt at replacing RetinaNet with FCOS in the DeepForest library.
Uses the same NeonTreeEvaluation dataset for benchmarking.
Trained on an NVIDIA T4 using GCP:
| Library | Precision | Recall |
|---|---|---|
| CrowNNs | 0.64 | 0.43 |
| DeepForest (SOTA) | 0.66 | 0.79 |
| lidR package (2016) | 0.34 | 0.47 |
Make sure CUDA is installed.
Create a virtual environment using
python3 -m venv .venv
and install requirements.txt.
python3 -m train.main <args>
python3 -m evaluation.<script>.py