I want to build a Jewellery similarity application.
I have around 70k images.
The images will have the following characteristics:
- Type - Ring, Bracelet, Earring
- Material - Gold, Silver, Titanium
- Shape - Round, Heart,etc
- Color
Now this is how I believe the system will be -
I will use Resnet50. Will remove the last layer.
The last layer will be as per the 4 characteristics given above.
Am I in the right direction?
Another ask -
Also, for training the images, majority of them have the background white. But some have transparent. I shall convert the transparent ones to white. Right?
Another ask -
Since client wants to run everything on cloud. After training, the inferences’ second last layer embeddings will be saved in a Vector DB like Pinecone.
Am I in the right direction?