This jupyter notebook is to train a lightweight MobileNetV2 model for disaster classification using the Crisis Image Benchmark Datasets (CrisisIBD)
A MobileNetV2 is trained for Disaster Classification using the procedures as described in the paper.
The notable difference is I use weaker augmentations for my training.
Based on experiments, fine-tuning MobileNetV2 (pre-trained on ImageNet) is often less steady if the augmentations are too strong.
The benchmark values are extracted from Table 11 of the paper.
I only compare with the MobileNetV2 results in the paper, as I only intend to train a MobileNetV2 with performance as close as the one described in the paper.
| Model | Acc. | Prec. | Recall | F1 |
|---|---|---|---|---|
| MobileNetV2 (paper) | 0.785 | 0.781 | 0.785 | 0.782 |
| MobileNetV2 (mine) | 0.776 | 0.787 | 0.776 | 0.781 |
- Train a Multi-Task Model for Disaster Classification and Victim Detection
- Use other CNNs (such as EfficientNet)
Priority is given for first task, as it is part of my research project.
Feel free to contribute ya :)