Open vigneshgig opened 5 years ago
Your problem seems similar to issue #1. TextBoxes requires axis-aligned bounding boxes, but TextBoxes++ and SegLink require oriented bounding boxes. I was simply too lazy to implement the 'polygon' case for datasets containing only axis-aligned bounding boxes.
The implementation of the 'polygon' case in the corresponding GTUtility
is actually pretty straightforward. For the SVT dataset, you can take b18d09d8e61af26a1551ae49dd98be4768943b30 as an example.
from data_svt import GTUtility
gtu_train = GTUtility('data/SVT/', polygon=True)
gtu_test = GTUtility('data/SVT/', test=True, polygon=True)
If you find the time, pull requests are welcome :)
How can i use custom dataset which does not have orientation bounding box alignment, I using LabelImg tool to create custom pascal format dataset. Thanks for the reply
@sivatejachinnam Take data_voc.py
, remove the conversion stuff, fix the class names, add the text
attribute and add the 'polygon' case as in the SVT example... should not be that hard...
How do I use LabelImg tool to create custom dataset? For example, I have an image below needs to be labeled, then how to do the labeling? Should I give all the text ("Campus" and "Shop" in this example) the same class name "text" (as shown in the screenshot below), or the exact letters in the text (i.e. "Campus" and "Shop")?
Thanks!
i loaded the dataset using below code
then i runned the code
and finally
but i get this error
/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/gradients_impl.py:110: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory. "Converting sparse IndexedSlices to a dense Tensor of unknown shape. " Epoch 1/100
ValueError Traceback (most recent call last)