Re-distributing Biased Pseudo Labels for Semi-supervised Semantic Segmentation: A Baseline Investigation

Abstract

While self-training has advanced semi-supervised semantic segmentation, it severely suffers from the longtailed class distribution on real-world semantic segmentation datasets that make the pseudo-labeled data bias toward majority classes. In this paper, we present a simple and yet effective Distribution Alignment and Random Sampling (DARS) method to produce unbiased pseudo labels that match the true class distribution estimated from the labeled data. Besides, we also contribute a progressive data augmentation and labeling strategy to facilitate model training with pseudo-labeled data. Experiments on both Cityscapes and PASCAL VOC 2012 datasets demonstrate the effectiveness of our approach. Albeit simple, our method performs favorably in comparison with stateof-the-art approaches. Code will be available at https: //github.com/CVMI-Lab/DARS.

Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation

“…Table 3 shows that our approach outperforms the SOTA methods for each architec- ture and backbone settings. For PSPNet, our approach outperforms DARS [18] by 0.7% mIoU and CCT [34] by 5.19%. In the experiments, our approach outperforms other SOTA approaches by a large gap.…”

“…We load the ImageNet pre-trained checkpoint, and the segmentation heads are initialized randomly. Following previous papers [9,18,34], we utilise the following polynomial learning-rate decay: (1 − iter max iter ) 0.9 . We also test our method on PSPNet [22,34] to show the generalization of our approach.…”

“…On Pascal VOC 2012, we crop images to 512 × 512 pixels for DeeplabV3+, train for 80 epochs with base learning rate set to 0.01, using batch size 32, for both labelled and unlabelled images, following [9]. For PSPNet, we follow [18] and crop images to 321 × 321 pixels and use batch size 8. On Cityscapes, due to hardware limitation, we crop images to 712 × 712 pixels, train for 450 epochs with base learning rate set to 0.0038 with batch size 8 for both architectures.…”

“…Even though L2 loss is known to be robust, which is advantageous when dealing with the noisy predictions produced by consistencybased methods, it is also known to have poor converge and to possibly lead to vanishing gradients. Given the reliability of the segmentation predictions produced by our extended MT model, we instead use the more effective cross entropy (CE) loss, constrained to be computed at regions of high-confidence segmentation results, represented by c(ω) in (3), following the strategy applied in self-training approaches [18,39,41].…”