thanhhau097 / chargrid2d Goto Github PK

Hello @thanhhau097 ,
I tried to train your implementation with custom data, and I'm getting what I consider high values of IOU (80-90%) in validation, but when I try to run the tests I rarely even get a single result (accuracy < 1% on 4 different labels).
Do you know what could be causing this?
Thank you for open sourcing this.
Best,
Mirko

Hi,

Can you please elaborate on the approach used for labeling data and creating the ground truth segmentation mask?

Hello,

I was trying to test the trained model with new data. Could you please help me with the pipeline involved in doing so?
I ran the generate_mask.py with the test data and then ran test.py, but I was getting a size mismatch error.

Kindly help. Thanks!

Hi, thanks for the implementation. I have a confusion about the reason authors came up with the bounding box regression decoder part. Is it for instance segmentation? As much as I understand, it helps to distinguish instances from the same classes. Is this correct? Also, I have a custom dataset which contains just the detected words by OCR and labels for each word. It is like a Named Entity Recognition (NER) dataset. I do not have interested regions annotated like items, company name etc. I have only words, bounding box and label for each word. In this case, I have implemented only the semantic segmentation decoder. Since I am not interested in instance segmentation, I thought that I do not the bounding box regression decoder branch. Is this a correct assumption or does the bounding box regression decoder improves the accuracy?

Can you please provide one demo Labelling data

thanks for your work, i trained on SROIE with your code, the loss decreased from about 50000+(the first epoch) to 10+(the 99th epoch). Is that normal?

Hello,

Thank you for sharing the implementation. I am a bit confused about what all inputs are needed and where to store them.

Hello, I had a question about the license. What kind of license is this code under ?.

Hi,

i have some question about your project here. Is this an implementation of the paper:
https://arxiv.org/pdf/1809.08799v1.pdf
Chargrid: Towards Understanding 2D Documents

Im working on recreating this paper and have some problems.
Thanks in advance

According to Figure 2 in the paper, the author did not apply downsampling in the first a block, but I found that in chargrid2d/model.py downsampling was applied in self.block1, thus dividing the height/width of all feature maps by factor 2. I guess that this adaption may help with the shape mismatch problem in the subsequent lateral connection step, as the descriptions of this part in the paper is ambiguous. Do you have any idea on concatenating two feature maps in different shapes if I did not apply downsampling at the beginning?

Hi, thanks for sharing you chargrid code. I read the chargrid paper but I'm a bit confused about the anchor box representation and the corresponding bounding box representation. How do you generate the ground truths for the bounding box regression decoder?

Thanks for putting up the implementation of Chargrid. Kindly explain the steps to execute the code if I have set of invoice images with annotation of relevant areas which need to be extracted. I don't have the exact labels of those designated areas in a json file.

Hello,

Can you please share the code and results for the SROIE Dataset?