DiCaP: Distribution-Calibrated Pseudo-labeling for Semi-Supervised Multi-Label Learning

TL;DR

DiCaP addresses noisy pseudo-labels in semi-supervised multi-label learning by weighting pseudo-labels according to their estimated correctness. The authors prove that the optimal weight is the posterior correctness given the predicted confidence and implement a practical confidence-bin estimation, coupled with a dual-thresholding mechanism and a class-wise contrastive loss for uncertain samples. A final fine-tuning stage ensures full utilization of labeled data. Empirical results across COCO, VOC, NUS-WIDE, and AWA show consistent improvements over state-of-the-art methods, especially under severe label scarcity.

Abstract

Semi-supervised multi-label learning (SSMLL) aims to address the challenge of limited labeled data in multi-label learning (MLL) by leveraging unlabeled data to improve the model's performance. While pseudo-labeling has become a dominant strategy in SSMLL, most existing methods assign equal weights to all pseudo-labels regardless of their quality, which can amplify the impact of noisy or uncertain predictions and degrade the overall performance. In this paper, we theoretically verify that the optimal weight for a pseudo-label should reflect its correctness likelihood. Empirically, we observe that on the same dataset, the correctness likelihood distribution of unlabeled data remains stable, even as the number of labeled training samples varies. Building on this insight, we propose Distribution-Calibrated Pseudo-labeling (DiCaP), a correctness-aware framework that estimates posterior precision to calibrate pseudo-label weights. We further introduce a dual-thresholding mechanism to separate confident and ambiguous regions: confident samples are pseudo-labeled and weighted accordingly, while ambiguous ones are explored by unsupervised contrastive learning. Experiments conducted on multiple benchmark datasets verify that our method achieves consistent improvements, surpassing state-of-the-art methods by up to 4.27%.

Figures (3)

• Figure 1: (a) Correctness likelihood distributions of labeled and unlabeled data after warm-up under 5% labeled setting on COCO. (b) Correctness likelihood distributions of unlabeled data across models trained with varying amounts of labeled data on COCO. (c) Estimated and true distributions for unlabeled data under 5% labeled setting on VOC. (d) Performance comparison between uniform and correctness-weighted pseudo-labeling on COCO.
• Figure 2: Illustration of estimation set construction and relationships among $\mathcal{D}_l$, $\mathcal{D}_u$, $\mathcal{D}_{\text{sup}}$, $\mathcal{D}_{\text{est}}$, and $\mathcal{D}_{\text{unsup}}$.
• Figure 3: Comparison between estimated and optimal correctness likelihood distributions at early and late training stages on COCO, NUS and AWA under 5% labeled setting.