Question about data quantity discrepancy in description/conversation samples

[Lay-du]

Hi, this is a great job, but I have some questions I would like to ask you.

When constructing the description and conversation data pipelines based on the provided datasets, I noticed an interesting phenomenon that I'd like to understand better.

According to the implementation logic (as I understand it), we should be creating one description/conversation sample per image in the union set of RefCOCO/RefCOCO+/RefCOCOg images. However, when examining the publicly released conversation and description instruction-tuning data, I observed that:

The number of conversation/description samples exceeds

The total number of unique images in the combined RefCOCO/RefCOCO+/RefCOCOg datasets

Could you help clarify what causes this discrepancy? Some possibilities I considered but would like to confirm:

Are multiple descriptions/conversations being generated for a single image?

Does the dataset incorporate additional images beyond the standard RefCOCO series?

Is there data augmentation or synthetic data generation involved?

This understanding would greatly help in aligning the data construction process with the intended design. Thank you for your insights!