In this repository, I am sharing how to do tree species identification using only an image of a bark. Please note that in this project, I didn't follow how the original authors splitted the data. For more details about data splitting or any other details, please refer to the original paper.
The dataset contains 23 different species of trees found in parks and forests near Quebec City, Canada. It contains more than 23,000 images. The dataset were splitted as follow:
| Dataset | Train | Validation | Test |
|---|---|---|---|
| BarkNet 1.0 | 15,057 | 3,765 | 4,706 |
For this project, we have tested 4 models: MobileNet V2, MobileNet V3, EfficientNet-B0 and MobileViT.
Here are the obtained results:
| Model | TOP-1 Accuracy | TOP-5 Accuracy |
|---|---|---|
| MobileNet V2 | 78.83% | 97.69% |
| MobileNet V3 | 80.42% | 97.84% |
| EfficientNet-B0 | 80.69% | 97.54% |
| MobileViT | 79.35% | 97.92% |
Run the project in background
nohup python main.py --model_name MobileNet2 > mobilenet2_nohup.out &
nohup python main.py --model_name MobileNet3 > mobilenet3_nohup.out &
nohup python main.py --model_name EfficientNet-B0 > efficientnet_b0_nohup.out &
nohup python main.py --model_name MobileViT > mobilevit_nohup.out &
While the models are training, you can run the following command to visualize the training loss and accuracy using tensorboard
tensorboard --logdir=utils/logs/runs/ --bind_all
Here how it looks like to visualize the training loss and accuracy
