search

Insta-Cook / native-cv

0 stars 0 forks source link

A few ideas to enhance the model #5

Open mohasarc opened 2 years ago

mohasarc commented 2 years ago

So we've talked before about using background subtraction to enhance the detection.

Aside from that, I have an idea regarding the dataset, what do you think about synthesizing fridge images. basically, we can do the following:

This way we can efficiently generate thousands of images that reflect reality to some extent.

@arnisafazla please let me know what do you think about his idea.

mohasarc commented 2 years ago

I think we'll need background subtraction for the scraping and image collection step. The reason is that most of the collected images will probably have backgrounds, but we need the fruit with transparent background to be able to place it in the 3d model.

mohasarc commented 2 years ago

If that succeeds, we can change the setting and generate more images. We can, for example, place the items on a shelf model. or inside a plastic bag model.

mohasarc commented 2 years ago

Links to some fridge models we can use:

A whole kitchen model:

arnisafazla commented 2 years ago

Synthesizing images is a good idea, our supervisor said we can try that as well. Only problem is we will either need to figure out how annotation files work and construct those automatically as well, or use an annotation tool's API to annotate automatically.

mohasarc commented 2 years ago

I see, let's research that, I feel like there should be an easy way to automate that.