Any idea why this image was missed from extraction?

[LI3ARA]

I have run code to extract this pdf. Though for several times I did not see the image has not been extracted.
Visualized for better understanding with the current parts that are extracted.

What could be the reason for this, and what might work for this kind of problem.

[joshus00]

I have something similar where I can extract captions for all images but not annotations.

I was following this page
https://docling-project.github.io/docling/examples/pictures_description/

If I run this locally and I print picture 5, I just get the caption and not the annotation (the LLM generated description) as so

Picture #/pictures/4

Caption
Figure 5: Conversion time in seconds per page on our dataset in three scenarios, across all assets and system configurations. Lower bars are better. The configuration includes OCR and table structure recognition ( fast table option on Docling and MinerU, hi res in unstructured, as shown in table 1).

Annotations (HuggingFaceTB/SmolVLM-256M-Instruct)
The image is a bar chart titled "see/page." The x-axis represents the number of pages, while the y-axis represents the page count. The chart is divided into four sections: "Docling," "Marker," "Mineru," and "Unstructured." Each section has a different color and a corresponding number of pages. - Docling: The chart has a blue bar for "Docling." This bar is positioned at the left side of the chart. - Marker: The chart has an orange bar for "Marker." This bar is positioned at the top of the chart. - Mineru: The chart has a green bar for "Mineru." This bar is positioned at the bottom of the chart. - Unstructured: The chart has a black bar for "Unstructured." This bar is positioned at the right side of the chart. Each section has a different color and a corresponding number

Picture #/pictures/5

Caption
Figure 6: Ecosystem of Docling integrations contributed by the Docling team or the broader community. Docling is already used for RAG, model fine-tuning, large-scale datasets creation, information extraction and agentic workflows.

Note the missing annotation for picture 5 eventhough it clearly extracted it as an image.

I have been trying to figure out why it works for some and not others.

Any help would be appreciated!