Image classification household tools

Part 1 (2019)
1

I’ve just finished lesson 1. I’m trying to build household tools classifier. By now, I have 10 categories (~100 images per category). I’ve downloaded them from google search, added some pictures from online markets where people do sell their tools (to get more valid data) and manually deleted some outliers. By now my model has accuracy ~92-94% (ResNet-50, 12 epochs). But I suppose even worse accuracy for real images.

image
image.png463×552 21.9 KB

By experementing a little bit, I found out, that the main problem is in my dataset. When you download images from google, it may occur so, that an image you get is just “object you need placed on white background”.

Therefore, here are the questions:

  1. Are there some techniques that can help to replace white background with random background from our sets of backgrounds provided ?
  2. Do you know some good household tools datasets ?
  3. Should I leave this model as it is, deploy it, and start to collect user data as fast as I can ?

Thank you !

1 Like
2

this is an interesting idea. using maybe opencv contour you can identify object, remove them and place them on a random background like jungles to give it a wider set of training data. It can be tricky using opencv to accomplish this I have tried object extraction and failed multiple times, but it is a good idea.