← All posts

WSSS

Contract WSSS research · 53.31% COCO-Val mIoU · +1.5 pp over WeCLIP+

Built with PyTorch · CLIP · DINOv2 · Semantic segmentation · COCO-Val 2014

Bar chart comparing WSSS performance of WeCLIP+ at 51.8 percent mIoU and the refined method at 53.31 percent mIoU on all 40,137 COCO-Val 2014 images
COCO-Val 2014 · all 40,137 validation images · mIoU

WHY

A client commissioned a study to improve WeCLIP+, then the state of the art, using only image-level labels. The pseudo-masks used for training accumulated wrong pixels at object boundaries and in the background, and those errors were amplified during self-training. The research question was whether only the unreliable pixels could be identified and repaired instead of rebuilding every mask.

COCO-Val 2014 40,137 images

Evaluation over the complete validation set

Refined WSSS 53.31% mIoU

Final semantic-segmentation performance

Previous SOTA +1.5 pp

Compared with WeCLIP+ at 51.8% mIoU

HOW

From weak labels to reliable pixel supervision

The model preserves strong pretrained representations and spends learning capacity on the uncertain parts of each generated mask.

  1. 01 Frozen representations

    Extract complementary visual evidence with CLIP and DINOv2 without updating either backbone.

  2. 02 Reliability map

    Measure where the two representations support the same pixel assignment and where they disagree.

  3. 03 Pixel repair

    Replace only the unreliable regions instead of regenerating the entire pseudo-mask.

  4. 04 Segmentation training

    Train the feature-fusion heads and segmentation decoder using the repaired masks as supervision.

RESULT

COCO-Val comparison

MethodSupervisionmIoU
WeCLIP+Image-level labels51.80%
Refined pseudo-label methodImage-level labels53.31%

Both figures use the full COCO-Val 2014 evaluation set.

Weakly supervised segmentation architecture using frozen CLIP and DINOv2 representations, selective pseudo-mask repair, and self-training
Frozen visual representations, reliability mapping, selective pixel repair, and retraining

Contribution

  • Designed the pixel-reliability and selective-repair strategy around frozen CLIP and DINOv2 representations.
  • Implemented the refinement and retraining pipeline, then evaluated the final model over the complete COCO-Val 2014 set.

Evidence

Source code Training, refinement, and evaluation implementation