Closed kkissmart closed 3 years ago
The dataset which we used is smaller and the pre-trained model is also not production ready.
So there would be many scenarios where the model would fail. If your table looks similar to what we have trained on then there should not be any problem. If not, then you need to fine-tune the model for your own data.
@DevashishPrasad Thanks for your fast response
I am using your demo png, but I can't reproduce your result.
when I load this config, the mmdetection code said no 'num_stages' key recognizable, can you point me where this variable is used in the cascade rnn model in mmdetection codebase?
So i removed this num_stages=3. then the model can be loaded, but the result is very awful. Do I miss anything?
Thanks a lot!
this is what I run and what I got, super far away from what was claimed in the paper.
https://drive.google.com/file/d/1NvUFJUQGlsPqTitML-gw-Uqec6JQWbpE/view?usp=sharing
Please let me know if anything is missing
The Demo.png was made manually (using photo editing) just to provide an illustration of the idea. Those are not the actual results of cascadetabnet model.
You can find the actual results on more images in our CVPR paper.
But the results that you are getting on Demo.png are definetly poor and cascadetabnet would do far more better. Most probably you are missing something.
I used cascade_mask_rcnn_hrnetv2p_w32_20e.py as config and epoch_36.pth as model
from mmdet.apis import inference_detector, init_detector, show_result_pyplot
Choose to use a config and initialize the detector
config = 'configs/cascade_table/cascade_mask_rcnn_hrnetv2p_w32_20e.py'
Setup a checkpoint file to load
checkpoint = '/home/model/epoch_36.pth'
initialize the detector
model = init_detector(config, checkpoint, device='cuda:0') img = "/home/code/CascadeTabNet/Demo/demo.png" result = inference_detector(model, img) show_result_pyplot(model, img, result, score_thr=0.85)
the result is a mess. It has tons of tables but none of them are correct.