Weakly-Supervised Semantic Segmentation Pseudo Masks for COCO and VOC

Preprocessed auxiliary files for the ModuSeg project support weakly-supervised semantic segmentation. The dataset includes CorrCLIP-generated pseudo masks, VOC-style segmentation annotations for COCO2014, and augmented annotations for VOC2012 SBD, along with image-level labels in JSON format. Author QZing007 uploaded the collection to Hugging Face in April 2026.

Use Cases

• Training segmentation models using the provided CorrCLIP pseudo masks as weak supervision targets.
• Benchmasking model performance against the VOC-style SegmentationClass annotations for COCO2014.
• Utilizing the image-level labels JSON file for multi-label image classification tasks or as input for other pseudo-label generators.
• Augmenting the VOC2012 SBD dataset with the provided additional segmentation annotations.
• Comparing the quality of CorrCLIP-generated pseudo masks against traditional fully-supervised annotations like those in SegmentationClass.

Strengths

• Includes annotations for two major benchmarks: COCO2014 and VOC2012 SBD.
• Provides multiple annotation types: pseudo masks, ground-truth segmentation, and image-level labels.
• Dataset was updated on Hugging Face in April 2026.

Limitations

• Specific row counts, file sizes, and total number of images are unknown.
• The quality and accuracy of the CorrCLIP pseudo masks are not quantified in the provided description.
• License terms for use and redistribution are unspecified.

Provenance

Source

Annotations derived from COCO2014 and Pascal VOC 2012 datasets, with pseudo masks generated by the CorrCLIP model.

Collection Method

Pseudo masks were generated by the CorrCLIP model from image-level labels; other annotations are preprocessed from existing segmentation datasets.

Freshness

Last updated on the platform in April 2026.