I was doing something with the downloaded PDF image of a plat of the subdivision in which I live that also included real estate related documentation etc. This is a picture of documents that were probably on microfilm at one time. I shared it from Dropbox to recognize with Seeing AI. All the text information came out great with Seeing AI, along with a bunch of lines, and it also summed up well what the PDF was when I asked it to describe the document. In the middle of that, I asked if the document had any colors, then if any characters had color or were all black and white. Seeing AI said it couldn't find that information in the document both times. Later on, I tried a second time with different approaches to getting information about whether it was color or black and white, no luck.
I asked about the colors because I'm planning on printing the PDF, and I was wondering if it would exorcise the color ink cartridge in my printer to keep it from drying/clogging, LOL, so it wasn't really of any importance.
By OldBear, 9 December, 2024
Forum
iOS and iPadOS
Comments
scien discriptions should…
scien discriptions should work well, my understandingis that the llm is fed the text ocr extracts, not the actual image
Probably
That would make sense. I was just recognizing it from the share sheet in Dropbox. I will have to figure out how to open that PDF file from the Scene Description in the Seeing AI app. It seems like I've done that before with a PDF, but I only seem to have the option to open pictures from the camera roll. If I get back to it, I'll try to figure it out.
Update
I got back around to this and did look at the image of the plat with Seeing AI's scene description. The only way I found to do it was to extract the image from the PDF, save it to my camera role, then go into Seeing AI.
It said the map and text were in black and white. However, that raised another question. Do printers use the color ink cartridge when printing black? It turns out my printer does, so I probably don't have to worry about having to find something with color to print out every week to keep the ink flowing.
AI and text atributes
Hi OldBear,
AI is bad at identifying text atributes. A few months ago, I needed to know which words were bolded in a text, so I decided to use AccessAI, because I knew I could get a human to verify it to make sure it was accurate. It was only a few lines. This was an image of a document. It told me which words were bolded, but when I verified it with a human, it was totally wrong. It felt like it had just guessed random words from the text instead of actually processing the text and correctly identifying the actual words in bold. This was a few months ago though, but I still wouldn't trust AI to accurately identify text atributes like colours, fonts and formatting info.
Thank You
Hi, this is Saqib from the Seeing AI team.
As others suggested, indeed when asking questions about the document we only analyze the text and the layout of the document. At present this does not retain color information.
I hadn't heard of this need before, but will add it to our list - along with the desire to recognize PDF's as images not documents. It would be great from others who have similar scenarios they'd like us to consider supporting.
All the best, Saqib
Color Scenario
Thank you, Saqib.
I'm just thinking about this. There are documents that put certain text in red or other colors to distinguish it from the main text, a warning, for example. The change in font or attribute, as Tara points out, can also convey important information that someone might be interested in. Our screen readers have this information available when we're reading accessible documents or web pages. Sometimes PDF documents also have photos embedded in them, or in the text, so directly using them in the scene description channel might save a few steps one would need to take in a separate app.
re Saqib
Hi Saqib,
Thank you for your reply. Certain professions such as translation, proofreading and teaching require accurate knowledge of text atributes. I'm an EFL teacher online, and I sometimes have to know whether things are in bold or italics etc.. I used to do proofreading online, and once I was told to only correct the text in yellow. There are lots of instances where proofreaders need to know about text atributes. Depending on the environment, for example, Windows in conjunction with Word, you can get your screen reader to announce details like that, however in Google Docs, this feature doesn't work, and on iOS I have no idea what works and what doesn't when it comes to text atributes and VoiceOver. If you're dealing with a PDF image, you've got no hope of knowing the formatting information. In my experience, doing an OCR usually doesn't render font and colours etc.. It would be really great if AI could accurately capture and describe this type of information.