Here I have approached this problem by leveraging the semantic relationships between the embedding vectors of the animals of seen and unseen classes and linking them to visual features.
Model-1(Model-1 (2).ipynb) and Model-2(Model-2.ipynb) are the same ones as discussed in the report.
vlg final.docx is the Performance Report which I have also uploaded on Google Drive.
Note:
- These are only the beautified version of notebooks of the two best models that I came up with.
- Many different pretrained models were used in experimenting (ResNet, EfficientNet etc.) but I used MobileNetV2 mostly because it is more efficient even though more accuracy could have been achieved using deeper models with more parameters.