PointAugment and DGCNN tree species classification for 3DForEcoTech Tr3D Species Benchmark. Model achieved the highest F1 score of 0.767 during the training/validation phase.
| Model | Description | Reference |
|---|---|---|
| PointAugment | Adaptation of PointAugment model for tree species point cloud augmentations | (Li et al., 2020) |
| DGCNN | Adaptation of Dynamic Graph Covolutional Neural Network (DGCNN) model for tree species classification on point clouds | (Wang et al., 2019) |
| Folder | File | Description |
|---|---|---|
| root | species_classes.csv | csv of species and associated class number |
| root | main.py | Main script to run the model |
| augment | augmentor.py | The augmentor (generator) model |
| checkpoints/dgcnn_pointaugment_4096 | f1.png | Image of the validation and training F1 scores |
| checkpoints/dgcnn_pointaugment_4096 | loss_f1.csv | csv of the augmentor losses, classifier losses, and F1 scores |
| checkpoints/dgcnn_pointaugment_4096 | run.log | Run log of printed outputs |
| checkpoints/dgcnn_pointaugment_4096/models | best_model.t7 | Pytorch model weights of the best run |
| checkpoints/dgcnn_pointaugment_4096/output | confusion_matrix.png | Image of confusion matrix of best model |
| checkpoints/dgcnn_pointaugment_4096/output | output.csv | csv of true and predicted classes |
| common | loss_utils.py | The loss fucntions for the adapted models |
| models | dgcnn.py | Pytorch Implementation of DGCNN |
| utils | augmentation.py | A script that performs the manual augmentations on point clouds |
| utils | resample_point_clouds.py | A script that performs resampling of point clouds (current methods are fps and cluster fps) |
| utils | send_telegram.py | Functions that send telegram messages + photos |
| utils | tools.py | A script of useful functions |
| utils | train.py | A script that defines the training/validation/testing process |