comparing ICDAR2013 results

[SHaiHosh]

Hi Since the result from TextBoxes++ is quadrangle, how did you compare to ICDAR2013? This is how I did it:

    x1 = result[0] # x is column    LEFT
    y1 = result[1] # y is row       TOP
    x2 = result[2] #                RIGHT
    y2 = result[3] #                 TOP
    x3 = result[4] #                RIGHT
    y3 = result[5] #                BOTTOM
    x4 = result[6] #                LEFT
    y4 = result[7] #                BOTTOMleft = int(.5*(x1+x4))
    rt = int(.5*(x2+x3))
    top = int(.5*(y2+y1))
    bot = int(.5*(y3+y4))

Am I correct?

[MhLiao]

You can just use (xmin, ymin, xmax, ymax).

[SHaiHosh]

Thank you for the advice

[SHaiHosh]

@MhLiao well I implemented this rectangle solution i ran the network with 1024 scale (and 1024 input in deploy.prototxt) but failed to reproduce the results reported in the article. see below:

Calculated!{"recall": 0.5371689497716894, "precision": 0.8055118110236225, "hmean": 0.644525409450198

[gittigxuy]

@SHaiHosh ,could you please write a blog to show how did you do to use Textbox++ and ICDAR2013

[MhLiao]

@SHaiHosh The provided model(for icdar2015) is only trained with the ICDAR 2015 dataset, while without the ICDAR 2013 dataset. To evaluate the performance on ICDAR 2013, you should train the network with ICDAR 2013 dataset based on the model pre-trained on SynthText.