Some clarification on the models

[Frank071]

I am very interested and grateful for your project which, after a slight hickup due to an unwilling numpy, works great!

My current implementation heavily relied on the few examples I found, but I am not sure if this is the optimal approach... One of the things that is unclear to me is how to choose the yolo9 and cct model. I found the benchmarks, but that does not point me in the direction of neither using yolo...256... or yolo...608..., nor the choice between v1 or v2, s or xs in cct...

It would be great if the documentation could shed a light into this...

[ankandrew]

Hi, I'm glad you find my project useful!

For the detector models available, most use the same architecture (the t, or tiny, variant of YOLOv9), but the number (256, 384, 512, etc.) changes. This number is the input size the model sees. In general, a bigger resolution helps detect smaller plates, but it also takes more time to process. If you are doing ALPR on a camera that is very close to the plate, or in other words the plate appears large in the image, then I would use the 256 variant, for example.

Regarding the recognition modules of fast-plate-ocr, v2 have region recognition for 65~ countries and have been trained with 3x more data. I wouldn't use v1 models given v2 are out. The S model has more params (compared to XS) so it should be more accurate. You can see some other diff between v1 and v2 here in v1.1.0 changelog.

[Frank071]

Thansk for the reply - as I found no (other) means to communicate, I am misusing this "issue" to at least inform you about how I am using your product. For a small festival I need a phone app with which we can identify people (cars) that have access. After havikng little success with tesseract and none with easyocr (could not get it installed), I found this gem.
Coming from tesseract, I have a web page that uses javascript to fetch and upscale (2.5 times) part of the camera image where the plate is. The image size depends on the device, for mine it is 1600w x 480h and the plate is somwhere in that frame. This image is sent to the server which then uses fast-alpr to extract the license plate. I used xs/384 before and am now testing with x/512. Not sure if I will see a difference...
In the tesseract set up I made the image client-side black and white only, which made it a lot smaller - but for fast-alpr I gathered it needed the coloured image, which is 150 kB. As the webapp can be set to continuous once-per-sec scan, this is a bit much, given data bundles etc. So I am now trying out a non-upscaled version (for my device 640x200ish, 20~50 kB), but am open to suggestions.

(and sorry for clobbering your issue list, but this is really intended to let you know it is used ;-) )