famerL/sam_dataset

This dataset is prepared for training SAM (Segment Anything Model) to perform zero-shot transfer.

The dataset consists of 7 sub-datasets, all of which are uploaded here. The download links and processing procedures for each sub-dataset are described in their respective README files, and the processing code is included within each corresponding ZIP folder.

It has been discovered that merging the seven subsets into a single set of images and ground truth masks results in inconsistent image formats, specifically a mix of PNG, JPG, JPEG, and BMP files.

Two approaches are being considered. The first approach involved using a train/val.txt file to accommodate all formats, with corresponding modifications made to train.py and utils/dataloader.py. However, this attempt failed as the train.py execution became stuck at the --- create training dataloader --- step.

The second approach is to convert all ground truth masks to PNG format and all images to JPG format. This is currently being attempted.