Open Duv54 opened 4 years ago
To increase detection speed, just a couple ideas:
scale
to 1 for pipeline.Pipeline
and pass fixed size images, I expect you'll get a speed improvement.Thanks for your response.
I will take a look. I saw that in tensorflow v2 you have config.optimizer.set_experimental_options()
that could speed up the inference time.
Hey guys, in my image there is more than 500 words and detection takes average 10 seconds on colab GPU can you tell me it's normal? Also have you found any solution?
Hi @faustomorales, Currently, i want to reduce time and focusing on the detection part first. I seen the original craft repo is giving a prediction in 1.3 seconds on my image and the same image keras_ocr takes 3.4 seconds. When i look changes i found by change this line reduce time to 2.9 seconds https://github.com/clovaai/CRAFT-pytorch/blob/master/craft_utils.py#L45 segmap[np.logical_and(link_score, text_score)] = 0 to segmap[np.logical_and(link_score==1, text_score==0)] = 0 can you tell me why you change this?
Hello, I am working on scanned documents of approximately 2048 * ~ 2900 pixels which contain ~ 500/600 words. The detector works better than Tesseract's segmentation. However, detection time is important for me and it takes about 8 seconds for 1 frame with Keras_ocr while it takes 0.6s for Tesseract. How could I speed up the detection part? If I resize the images to a lower width and height, the processing time decreases enormously but the performance of the detector also decreases enormously until only detecting 3 boxes out of 500 present