search

RyersonMultimediaLab / DynaSeg

Official repository for DynaSeg: A Deep Dynamic Fusion Method for Unsupervised Image Segmentation. Includes clean code, datasets, and instructions to reproduce the results reported in our Image and Vision Computing journal paper.
MIT License
3 stars 0 forks source link

Training code for custom dataset #1

Open cvmlarun opened 3 months ago

cvmlarun commented 3 months ago

Training code for custom dataset

RyersonMultimediaLab commented 1 month ago

The script DynaSeg_COCO.py is built following MMDetection standards and is primarily used to generate a confusion matrix (accuracy, mIoU, and mAcc) for all classes, including "thing" and "stuff" categories, similar to DenseSiam and PiCIE.

However, if you are working with a custom dataset, it is recommended to use DynaSeg_SCF_BSD_VOC.py. This script can run independently, and you only need to specify the path to your dataset for it to function properly.

inna-kucher commented 3 weeks ago

I would like to take pre-trained resnet-18 , couple it with FPN and train on a custom dataset. I don't think DynaSeg_SCF_BSD_VOC.py would let me do it.

I cannot understand how the cluster labels are estimated, since the method to find the optimal clusters is not applied in this script ... Could you please let me know how the cluster labels are estimated ?

Thank you