TextBoxes_plusplus Bad detect result after training

I've trained this network with self defined dataset, configurations are listed below. num_train_image: 700, num_test_image: 300, pretrain_model: models/model_pre_train_syn.caffemodel, resize_height: 384, resize_width: 384, base_lr: 0.0001, batch_size: 50

After 1800 iterations, loss declined from 5.3 to 2.1, but the detection result looks strange. https://drive.google.com/open?id=1_T-Bi89nIU8Z7-xYc9w0Rj9BafpbfTz3 All detection boxes look like triangles and start from the left top corner.

My annotation files look like below, I don't think there are any problems with annotation files. Is there anyone with the same issue?

<annotation>
    <folder>train</folder>
    <filename>bCXpXXunYpLFXX.jpg</filename>
    <size>
        <width>800</width>
        <height>800</height>
        <depth>3</depth>
    </size>
    <object>
        <difficult>0</difficult>
        <content>THEREISAWAY</content>
        <name>text</name>
        <bndbox>
            <x1>407</x1>
            <y1>413</y1>
            <x2>407</x2>
            <y2>425</y2>
            <x3>598</x3>
            <y3>411</y3>
            <x4>600</x4>
            <y4>397</y4>
            <xmin>407</xmin>
            <ymin>397</ymin>
            <xmax>600</xmax>
            <ymax>425</ymax>
        </bndbox>
    </object>
    ....
</annotation>

Jun 07 '18 08:06 tianranwangcs

@BIOTONIC The vertexes of the quadrilaterals should be in clockwise.

Jun 07 '18 09:06 MhLiao

@MhLiao Thank you for your help. I've checked all my labels and changed the sequence of vertexes into clockwise but still have this problem. Any ideas?

Jun 08 '18 04:06 tianranwangcs

Hi, I have the same bad result. https://drive.google.com/file/d/1HurdKGzqSudUF3yhHPpNn4TGDvL0h0qd @BIOTONIC Are you fixed the error ?

Jul 31 '18 21:07 khalilo

which tool did you use for label image?

Oct 08 '18 03:10 MarStarck

which tool did you use for labeling images?

Did you find which tool was used?

Mar 01 '23 04:03 dahalsweekar