Question about Layout Location Coordinate Normalization.

[KimRass]

I don't understand the meaning of layout location coordinate normalization.

The input image resolution during pre-training is 1920 x 1600, but the coordinates are quantized into the range $[0, 1000]$.

Does this mean that any ground truth coordinates greater than 1000 are simply clipped to 1000?

[Veason-silverbullet]

@KimRass Coordinate bin normalization means all coordinates are quantized into a relative range of $[0, 1000]$.

Suppose a pixel of the image in the location $(x_1, y_1, x_2, y_2)$, the pre-training size H x W is 1920 x 1600, the normalized location is $\big(\frac{x_1}{W} \times 1000, \frac{y_1}{H} \times 1000, \frac{x_2}{W} \times 1000, \frac{y_2}{H} \times 1000\big)$.

For example, a pixel of the image in the location $(320, 192, 640, 960)$ with the pre-training resolution (H x W) = (1920 x 1600), its normalized location is $(200, 100, 400, 500)$.

Also, note that the pre-training and fine-tuning resolutions are different.

[KimRass]

I fully understood. Thank you very much!

[KimRass]

So in files such as 'ocr_0.json' coordinates are already quantized. Is this right?

[Veason-silverbullet]

So in files such as 'ocr_0.json' coordinates are already quantized. Is this right?

Yes.