Sinkhorn Output Perturbations: Structured Pseudo-Label Noise in Semi-Supervised Segmentation

Robert Mendel, David Rauber, Christoph Palm

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: semi-supervised, segmentaiton, optimal transport, strong-weak augmentations

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

In semi-supervised segmentation, the strong-weak augmentation scheme has gained significant traction.
Typically, a teacher model predicts a pseudo-label or consistency target from a weakly augmented image, while the student is tasked with matching the prediction when given a strong augmentation.
However, this approach, popularized in self-supervised learning, is constrained by the model's current state.
Even though the approach has led to state-of-the-art improvements as part of various algorithms, the inherent limitation, being confined to what the teacher model can predict, remains.

In Sinkhorn Output Perturbations, we introduce an algorithm that adds structured pseudo-label noise to the training, extending the strong-weak scheme to perturbations of the output beyond just input and feature perturbations.
Our strategy softens the inherent limitations of the student-teacher methodologies by constructing noisy yet plausible pseudo-labels.
Sinkhorn Output Perturbations impose no specific architectural requirements and can be integrated into any segmentation model and combined with other semi-supervised strategies.
Our method achieves state-of-the-art results on Cityscapes and presents competitive performance on Pascal VOC 2012, further improved upon combining our with another recent algorithm.
The experiments also show the efficacy of the reallocation algorithm and provide further empirical insights into pseudo-label noise in semi-supervised segmentation.
Code is available at:

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7406