Bug

[Coxinha96]

请提供下述完整信息以便快速定位问题
(Please provide the following information to quickly locate the problem)

• 系统环境/System Environment: Windows
• 使用的是哪门语言的程序/Which programing language: Python
• 所使用语言相关版本信息/Version: 0.0.9-0.0.10
• OnnxRuntime版本/OnnxRuntime Version: Chinese
• 使用当前软件/库的版本/Use version: 0.0.9-0.0.10
• 可复现问题的demo和文件/Demo of reproducible problems:
• 完整报错/Complete Error Message:
• 可能的解决方案/Possible solutions:

This is the first time I'm reporting something here. I apologize if I did something wrong.

The problem is the following: when I do the extraction, everything goes "normally", as soon as it finishes, nothing happens. The program doesn't do the OCR. VideoSubFinder remains open in the Task Manager consuming CPU for a long time, consuming around 10% - 20% of the CPU.

I've waited for 2 hours, and nothing, the OCR doesn't happen. This has been happening since 0.0.8. Version 0.0.4 is the one I use the most, and it's the oldest version I have and it works quite well, but the OCR isn't as accurate as the new version.

What could it be? If I only do the manual OCR, it works. But I wish I could use the Extraction and the OCR, so I can do it with several files.

[SWHL]

Could you provide more information about the problem?
I need to be able to reproduce the problem in order to possibly solve it.

[Coxinha96]

Last edit:

Coming back here a few days later to update.

I ended up creating my own script to run the rapid_videocr library.

It's working perfectly now. Fast and very accurate.

Thank you for your attention!

Thank you for your reply!

I used ProcessMonitor to analyze the behavior of both VideoSubFinder and RapidVideOCR. I noticed that VideoSubFinder is receiving a command-line argument to generate the "TXTImages" folder.

What I observed is that before RapidVideOCR performs its OCR, VideoSubFinder is processing/creating TXTImages. In my opinion, this is not the ideal behavior, since it is generally recommended to run OCR on the "RGBImages" folder. This behavior does not occur in earlier versions of RapidVideOCR.

Ideally, the command to generate "TXTImages" should either be optional or customizable, so that the user can modify the command-line arguments passed to VideoSubFinder. That way, the configuration would be more flexible.

In my case, version 0.0.10 doesn't proceed to OCR after VideoSubFinder processes the TXTImages folder, which is also the most time-consuming part. Version 0.0.9 is slightly faster, even though it still generates TXTImages, but the delay is still noticeable.

Currently, I'm using version 0.0.4. I tried replacing the models with the ones from 0.0.10 (ch_PP-OCRv4_det_infer and ch_PP-OCRv4_rec_infer) and updated the config.yaml accordingly, but it doesn’t seem to work properly. The results are not as accurate as the manual OCR I perform using version 0.0.10.

---

Edit:
My apologies!
After replying earlier, I ran some additional tests and realized that version 0.0.4 also generates the TXTImages folder — however, it's much faster! I'm now confused about why the newer versions take so much longer to complete this step.

For reference, my system specs are:
CPU: AMD Ryzen 7 5700X
GPU: NVIDIA RTX 2060 Super

Also, I have another question:
Is it possible to use the GPU for both OCR and VideoSubFinder? When I run VideoSubFinder directly to analyze a video, I can enable GPU acceleration using the Open Video (FFMPEG) option.
Is there any way to make RapidVideOCR pass this command to VideoSubFinder, so that GPU acceleration is used during this process as well?