search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.38k stars 648 forks source link

ABCNet data preparation #102

Closed shuangyichen closed 4 years ago

shuangyichen commented 4 years ago

Thanks for your great work and sharing the code. I wonder what is the number of point to annotate the curved text area.

Yuliang-Liu commented 4 years ago

For ABCNet training, we do not need point annotation.

shuangyichen commented 4 years ago

txt format annotation looks like x y x y x y ...x y x y ||||text So what is the number of groups (x y) needed?

Yuliang-Liu commented 4 years ago

@shuangyichen As much as needed. For example, for Total-text, the number of points is 10 (5 upper boundary and 5 lower boundary; for CTW1500, the number is 14.

shuangyichen commented 4 years ago

@Yuliang-Liu But in generate_bezier_json.py, segs= [float(kkpart) for kkpart in parts[:16]] `. So this part is only for data with 8 points annonated?

shuangyichen commented 4 years ago

@Yuliang-Liu Actually, I use the pretrained model tt_attn_R_50.pth to inference my own data, and the results were good but I still want to improve it. I made my own dataset following the steps the other issue mentioned. After training, the inference results were even worse than the pre-trained model. So I guess maybe the problem is in the process of generating annotation json. Because the number of each group(x,y) is different. Could you plz give me some advice how to train or just locate my problem.

shuangyichen commented 4 years ago

In generate_bezier_json.py

print("Usage: python convert_to_detectron_json.py root_path phase split") print("For example: python convert_to_detectron_json.py data train 100200")

what dose split mean?