Image and Annotation File Structure for own Training

johaq commented 7 years ago

Hi, I may be blind and just missing the obvious but is there a tutorial explaining in what format images and their annotations have to be? Darknet says: fu.png in train list, provide fu.txt with content classId xmin ymin width height

Is there an equivalent for Darkflow?

johaq commented 7 years ago

Ok I found this:

<annotation>
    <folder>VOC2007</folder>
    <filename>000001.jpg</filename>
    <source>
        <database>The VOC2007 Database</database>
        <annotation>PASCAL VOC2007</annotation>
        <image>flickr</image>
        <flickrid>341012865</flickrid>
    </source>
    <owner>
        <flickrid>Fried Camels</flickrid>
        <name>Jinky the Fruit Bat</name>
    </owner>
    <size>
        <width>353</width>
        <height>500</height>
        <depth>3</depth>
    </size>
    <segmented>0</segmented>
    <object>
        <name>dog</name>
        <pose>Left</pose>
        <truncated>1</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>48</xmin>
            <ymin>240</ymin>
            <xmax>195</xmax>
            <ymax>371</ymax>
        </bndbox>
    </object>
    <object>
        <name>person</name>
        <pose>Left</pose>
        <truncated>1</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>8</xmin>
            <ymin>12</ymin>
            <xmax>352</xmax>
            <ymax>498</ymax>
        </bndbox>
    </object>
</annotation>

Are the tags owner and source necessary? And is the the content of folder a relative path? What does the attribute depth in tag size mean?

Matsam195 commented 7 years ago

From my own experiments, only the filename, size, segmented, and object tags are necessary for annotation files. As far as I know, depth is not important, so you can just fix it to 1.

jubjamie commented 7 years ago

I made a script today to convert my dataset to what is required for Darkflow and after analysing the code (and testing) this is all you need. I did have some trouble with filenames and regex errors so I had to just number all my files like the VOC dataset. See #284 and my answer there for a bit more info.

<?xml version="1.0"?>
<annontation>
<folder>images</folder>
<filename>10.jpg</filename>
<size>
<width>450</width>
<height>328</height>
</size>
<object>
<name>pig</name>
<bndbox>
<xmin>19</xmin>
<ymin>84</ymin>
<xmax>144</xmax>
<ymax>236</ymax>
</bndbox>
</object>
</annontation>

p.s. I don't actually think that folder is required but I didn't dig far enough to verify that so left it in. Happy to help further as i've spent all day trying to understand it myself. And finally have!

abagshaw commented 7 years ago

@jubjamie Pretty sure the folder is not required as you already specify it in the flags with --dataset to specify where your images are to be found.

jubjamie commented 7 years ago

Yeah I don't think it is either. Oh well it's in now. it's only 2 lines of code :p

@johaq you can have a look at this tool here to create VOC XML files. I've not tried it myself but it looks like it should do the job. I will be trying it out soon!

https://github.com/tzutalin/labelImg

johaq commented 7 years ago

@jubjamie Thanks a lot for the info! I also wrote a script that converts my current bounding box files into xmls. I appear to also have the filename error you described so i will rename my files.

I think it would be nice to be able to have images in different folders. Like class1/images class2/images etc. If I understood correctly I have to have all files in the same folder.

jubjamie commented 7 years ago

I believe that you do need them in the same folder however you may be able to re-write the filename to be within a directory and relative to the path set by --dataset. But you're annotation will need to change.

Also remember that your image must be annotated with boxes and those annotations can contain multiple classes for one single image. Having images separated by class in this instance is not the most sensible solution.

aayush-k commented 7 years ago

@johaq could we use your script for converting the bounding box files to xmls? Thank you so much!

johaq commented 7 years ago

@aayush-k https://github.com/CentralLabFacilities/object_recognition/blob/master/scripts/darknet_to_darkflow.py

This is the script I used to convert darknet bounding boxes to darkflow. If you use the debug flag it shows you the image region of the new bbs so you should be able to check if the conversion was successful.

Edit: of course you have to make some ajustments like image size and your class labels. You might also need to adapt either your file structure or the way the script parses yours.