RapidOCR vs Paddleocr - slower and worse ?

[zoldaten]

i have inference compared on this image -

and got this results:

here`s code to compare:

"""
from rapidocr_onnxruntime import RapidOCR
import time 
engine = RapidOCR()
#print(dir(engine))

st=time.time()
#img_path = 'tests/test_files/ch_en_num.jpg'
img_path = 'tests/test_files/3.png'
result, elapse = engine(img_path)
#print(result)
#print(elapse)
for idx in range(len(result)):
    res = result[idx]
    if res:
        print(res[1])
        
print('time: ',time.time()-st)

import time
from paddleocr import PaddleOCR,draw_ocr
ocr = PaddleOCR(use_angle_cls=False, rec_model="ch_PP-OCRv2_rec_infer_opt.nb",drop_score=0.9,\
                warmup=True,show_log=False,use_onnx=False, precision='int8',rec_algorithm='SVTR_LCNet', lang='en', \
                use_mp=True, total_process_num=3) # need to run only once to download and load model into memory

##from paddleocr.ppocr.utils.logging import get_logger as ppocr_get_logger
##ppocr_get_logger().setLevel(logging.ERROR)

st=time.time()
img_path = 'tests/test_files/3.png'

def ocr_me(frame):    
    result = ocr.ocr(frame, cls=False)
    for idx in range(len(result)):
        res = result[idx]
        if res:
            #print(res)
            r=[line[1][0] for line in res]
            print(r)

ocr_me(img_path)
print('time: ',time.time()-st)

looks strange...

[SWHL]

This is not surprising, as the two parameters and preprocessing strategies are not the same.

[SWHL]

I'm not sure what you want to do, but if you want to compare the difference in inference between the ONNX model and the original model in PaddleOCR, just use PaddleOCR and replace the ONNX model for comparison.

[zoldaten]

i think i could use RapidOCR inference and Paddle one to find out a winner.

[SWHL]

You can try both of them in your environment.

[zoldaten]

i have already done it as you can see.

[SWHL]

try more images