Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation

Abstract

Consistency learning using input image, feature, or network perturbations has shown remarkable results in semisupervised semantic segmentation, but this approach can be seriously affected by inaccurate predictions of unlabelled training images. There are two consequences of these inaccurate predictions: 1) the training based on the "strict" cross-entropy (CE) loss can easily overfit prediction mistakes, leading to confirmation bias; and 2) the perturbations applied to these inaccurate predictions will use potentially erroneous predictions as training signals, degrading consistency learning. In this paper, we address the prediction accuracy problem of consistency learning methods with novel extensions of the mean-teacher (MT) model, which include a new auxiliary teacher, and the replacement of MT's mean square error (MSE) by a stricter confidenceweighted cross-entropy (Conf-CE) loss. The accurate prediction by this model allows us to use a challenging combination of network, input data and feature perturbations to improve the consistency learning generalisation, where the feature perturbations consist of a new adversarial perturbation. Results on public benchmarks show that our approach achieves remarkable improvements over the previous SOTA methods in the field.

LaserMix for Semi-Supervised LiDAR Semantic Segmentation

“…In the context of semantic segmentation, CutMix-Seg [29] and PseudoSeg [44] apply perturbations on inputs and hope the decision boundary lies in the low-density region. CPS [13] and GCT [12] enforce consistency between two perturbed networks [45]. These perturbations [46,47,48], however, are either inapplicable or only yield sub-par results in 3D.…”

Translation Consistent Semi-supervised Segmentation for 3D Medical Images

“…The different views of consistency learning methods can be obtained via data augmentation [3] or from the outputs of differently initialized networks [2,4,5]. Mean teacher (MT) [4,[6][7][8][9] combines these two perturbations and averages the network parameters during training, yielding reliable pseudo-labels for the unlabelled data. However, the domain-specific transfer [3] of the teacher-student scheme can cause both networks to converge to a quite similar local minimum, reducing the network perturbation effectiveness.…”