Hello! I'm trying to create pipeline with several neural networks and one of them must be text-detection model. The sequence of tasks I want to implement is somthing like that: 1. human detection (pedestrian-detection-adas-0002 or person_detection_retail_0013 - the first works better for my task now). The cropped image based on the output of this node will be the input for other models running in parallel. 2. Next node is text detection model, that crops detected bbox and sends it to pytesseract module 3. Another model working with text detection in parallel (for example, trained yolo4tiny network for medical masks detection). The problem I faced is the compability of blobs created with different OpenVino versions. EAST model in the gen2-ocr example doesn't work with models compiled with OpenVino 2021.2 and I cannot find something useful in openvino pretrained models zoo (I tried to use horizontal_text_detection_0001 model for Myriad, but as for Intel support - It cannot be compiled for Myriad VPU due to some layers are not implemented). Can You advise me where to find pre-trained EAST model or something equal that can be compiled to required OpenVino version blob?

EAST combined with another models than used in gen2-ocr

asidsunrise

Hello! I'm trying to create pipeline with several neural networks and one of them must be text-detection model.
The sequence of tasks I want to implement is somthing like that:

human detection (pedestrian-detection-adas-0002 or person_detection_retail_0013 - the first works better for my task now). The cropped image based on the output of this node will be the input for other models running in parallel.
Next node is text detection model, that crops detected bbox and sends it to pytesseract module
Another model working with text detection in parallel (for example, trained yolo4tiny network for medical masks detection).

The problem I faced is the compability of blobs created with different OpenVino versions. EAST model in the gen2-ocr example doesn't work with models compiled with OpenVino 2021.2 and I cannot find something useful in openvino pretrained models zoo (I tried to use horizontal_text_detection_0001 model for Myriad, but as for Intel support - It cannot be compiled for Myriad VPU due to some layers are not implemented).

Can You advise me where to find pre-trained EAST model or something equal that can be compiled to required OpenVino version blob?

erik

Hello asidsunrise! Seems like an interesting challenge. What's the error when you try to run the text-recognition-0012 blob with other models? Aren't both versions of both models in the gen2-ocr demo 2021.2?

asidsunrise

Hello erik ! Text recognition blob works good in the gen2-ocr demo, but I don’t need it. I need lightweight Text detection model, because text detection blob provided in gen2-ocr doesn’t work with other blobs I compiled from intel pre-trained .bin models (for example human detection).

it would be fine to use this:
[(https://docs.openvinotoolkit.org/2021.2/omz_models_intel_horizontal_text_detection_0001_description_horizontal_text_detection_0001.html)]

but it cannot be compiled for myriad vpu. The two other Intel pre-trained text detection models (0003 and 0004) based on Pixellink seem to be slower and I can’t fully understand how to use their output format.

So I wonder if it is possible to find pre-trained text detection model like used in gen2-ocr (as I understand, it is EAST https://arxiv.org/pdf/1704.03155v2.pdf) in .bin IR format, that can be manually compiled to blob file?

asidsunrise

erik the attempt of using text detection model used in gen2-ocr produces error same to recent discuss:
https://discuss.luxonis.com/d/159-pipeline-createneuralnetwork-can-t-create-two-neural-networks

There isn’t such problem to train my own text detector, but I think there is no need of this if good pre-trained models exist. Thanks!

erik

@asidsunrise So why don't you just use the same NN blob text-detection.blob as in gen2_ocr demo? I believe it's compiled for 2021.2. As discussed in the link above, the problem is that you are trying to mix different blobs that don't have same openvino version. Since you want to use pedestrian detection retail 0013 I assume you copied the blob from the gen2-pedestrian-reidentification, however that blob is 2020.1 and thus you get the error mentioned. You should compile the pedestrian detection retail 0013 for 2021.2 and it should work fine. Here is our tool that can help a lot with compiling🙂

asidsunrise

Thanks! I didn't know about online conversion tool and it was very helpful for me. Windows installed Openvino
2021.2 didn't make compatible blob in my case and produced error like on screenshot.

Thanks to online conversion it's working now!

Brandon

Awesome, thanks for circling back!