- Scraped images from google using Javascript and Python Scripts
- Scraped about 1000 images
- This was a more diverse data as different garments of each category were used rather than using very similar images per class as in our previous Dataset.
For comparision purpose used the same classes as in first approach i.e Saree , Kurti and Shirt.
- Image Augmentation techniques were applied.
- Data was normalised and converted into tensor before passing into the model
ResNet-50
Training the last 2 layers i.e layer 4 and fc gave better results than training just the last layer. So 2 layer training was considered.
- Frozen Layers : conv1 , bn1 , relu, maxpool, layer1, layer2, layer3, avgpool
- Unfrozen Layers : fc. layer4
Test Loss: 0.393063
Test Accuracy: 83% (137/165)
