Datasets:

pixparse

/

idl-wds

like 176

Follow

Pixel Parsing 40

So I have been randomly sampling the image crops using bbox coordinates. I have noticed many problems:

1. Difference between the image crop and ground truth. I considered the possibility of filtering based on the confidence scores (0.99). But there are still some examples that can creep in.

2. Image crop coordinates are flipped.

3. The image crops contain text not present in the ground truth.

These are some of the major problems I have noticed with the dataset. This is a shame because the dataset is rich...