How to reproduce the same precision as in the paper?

likyoo / SegEarth-OV

SegEarth-OV: Towards Traning-Free Open-Vocabulary Segmentation for Remote Sensing Images

29 stars 0 forks source link

How to reproduce the same precision as in the paper? #5

Open KotlinWang opened 2 hours ago

KotlinWang commented 2 hours ago

I replicated the Potsdam data set according to the steps, and the results were much different from those in the paper. Did I make a mistake in any of my steps? 20241017110036

likyoo commented 2 hours ago

Please make sure that you have prepared the data correctly.

KotlinWang commented 2 hours ago

Yes, I built the dataset as required. 20241017111645

likyoo commented 2 hours ago

Could you show some of your cropped images and labels? Did you use a validation set for evaluation?

KotlinWang commented 2 hours ago

This is the result of running eval.py directly as prompted in README. 2_10_0_0_512_512_img 2_10_0_0_512_512_lable

KotlinWang commented 2 hours ago

Each cropped image appears to be a grayscale image.

likyoo commented 1 hour ago

I retested the script and successfully got the cropped RGB image. Make sure your raw data is correct.

likyoo commented 1 hour ago

KotlinWang commented 1 hour ago

Thank you very much! We used all zip files to build the dataset. Just use '2_Ortho_RGB.zip' and '5_Labels_all_noBoundary.zip' to get an RGB image.