dengdan / seglink

An Implementation of the seglink alogrithm in paper Detecting Oriented Text in Natural Images by Linking Segments
GNU General Public License v3.0
493 stars 177 forks source link

The f-measure of evaluation #24

Closed abc8350712 closed 6 years ago

abc8350712 commented 6 years ago

I used the command


python eval_seglink.py --checkpoint_path=./seglink/model.ckpt-136750  --dataset_name=icdar2015 --dataset_split_name=test --dataset_dir=./tf_records

to evaluate the model provided by you(seglink-384 model)

I change 'seg_conf_threshold' and 'link_conf_threshold' to 0.8 and 0.5 separately.

when i set the test image as 384x384, the result is Recall, Precision, Fmean = [0.48117587][0.72381693][0.57806695]

when I set the test image as 512x512, the result is Recall, Precision, Fmean = [0.61840743][0.78477693][0.69172925]

It doesn't match the result you provided. Is there anything i miss?

BowieHsu commented 6 years ago

if you want to test ICDAR results, the images size should be 1280 * 720 @abc8350712

abc8350712 commented 6 years ago

Thank You!!@BowieHsu