batch_recognition() batch_size affect confidence a lot

[vpckso]

I dropped text with low confidence to avoid detecting non-text (e.g. icon), so I want to have a reliable confidence value.

I have 22 slices to process, and checked confidence of a specific slice with different batch size.

batch size	confidence
1	0.81
2	0.88
3	0.92
4	0.89
5	0.88
6	0.87
7	0.86
8	0.83
12	0.81
24	0.67

Is it cause by padding and cannot avoid?

My workarounds:

• lower batch size (batch size > 8 does not speed up in my environment)
• re-run recognition on low confidence slices with low batch size