search

sergiomsilva / alpr-unconstrained

License Plate Detection and Recognition in Unconstrained Scenarios
Other
1.71k stars 604 forks source link

annotation-tool #53

Open G-YY opened 5 years ago

G-YY commented 5 years ago

Hi,I use the annotation-tool in your repo,but it did not work for me ,it always display "NO_LABEL",do you konw how to solve it?Thank you1 @sergiomsilva

Programmerwyl commented 5 years ago

This is an annotation tool If there is not label info in the .txt corresponding to the image, NO LABEL will be displayed This indicates that you need to enter the label information yourself image You can press space, then enter the information, and then press space to confirm the information After completion, pressing n will automatically generate the corresponding annotation information

Programmerwyl commented 5 years ago

For the label box of the license plate, it is also marked by the operation area of these keys, such as c, a, d, s To be honest, it doesn't work very well！ Often collapse

PhilipsKoshy commented 4 years ago

I used the annotation tool with hundreds of images, using 5 images each time. Using the mouse, place the cursor at the top left corner of the license plate, press c, a; then move the mouse to the top right corner of the license plate, press "a", move the mouse to bottom right, press "a", move finally to bottom left, press "a" for the last time; then press "space" bar; now you can enter the license number; press "Space" bar again; then press "n". This is the sequence I followed.

khawar-islam commented 4 years ago

Anyone work on annotation tool?. The tool mentioned in paper isn't working well

PhilipsKoshy commented 4 years ago

I have used the annotation tool mentioned in this repo for about 400 images. But, each time I ran the annotation tool commandline, I just used 5 images. So, I ran the commandline 80 times (80 iterations * 5 images/iteration = 400).

khawar-islam commented 4 years ago

After training on 400 images. Do you obtain a reasonable accuracy ?

PhilipsKoshy commented 4 years ago

Out of the three stages (Vehicle Detection (VD) + License Plate Detection (LPD) + OCR), the training mentioned in this repo is only for the LPD stage. VD and OCR uses the pre-trained models. Now, to answer your question, yes, the training improves the LPD (I had to train the LPD, as mine is for a different geographic region than what the repo LPD model is trained for). But, I abandoned this project as I did not get good results with two line plates, possibly due to poor OCR. My guess is the OCR training is not good enough for two line plates. It was able to recognize good-looking, simulated two-line plates, but it failed miserably with real two-line plates. I posted that question in this github and have not got any answers so far https://github.com/sergiomsilva/alpr-unconstrained/issues/72. Still looking for a good OCR stage...

laurentdebricon commented 4 years ago

@PhilipsKoshy can you tell me which project works for you ?

khawar-islam commented 4 years ago

@PhilipsKoshy Please email me. I am also working on the same project

thangdc94 commented 3 years ago

For labeling tool, I suggest using https://github.com/openalpr/plate_tagger It also has pre-build releases so you don't need to build from source. https://github.com/openalpr/plate_tagger/releases It's easy to use but after labeling you have to convert result to alpr-unconstrained format.

For two-line plates, I'm using some work-around. You can see my answer here https://github.com/sergiomsilva/alpr-unconstrained/issues/72#issuecomment-844953324

jackmleitch commented 3 years ago

Out of the three stages (Vehicle Detection (VD) + License Plate Detection (LPD) + OCR), the training mentioned in this repo is only for the LPD stage. VD and OCR uses the pre-trained models. Now, to answer your question, yes, the training improves the LPD (I had to train the LPD, as mine is for a different geographic region than what the repo LPD model is trained for). But, I abandoned this project as I did not get good results with two line plates, possibly due to poor OCR. My guess is the OCR training is not good enough for two line plates. It was able to recognize good-looking, simulated two-line plates, but it failed miserably with real two-line plates. I posted that question in this github and have not got any answers so far #72. Still looking for a good OCR stage...

I'm struggling to get improvements to the LPD with more training data. Could you please provide details to how you fine-tuned the model? Thanks in advance! @PhilipsKoshy