Deep learning-based 3D mesh segmentation methods typically rely on dense annotations and fully supervised training, which are costly and difficult to scale to diverse scenarios. This work proposes a light-weight labeling scheme and a corresponding weakly supervised learning framework that significantly reduces annotation cost while maintaining competitive performance. First, we project pixel-level labels from 2D rendered images back onto mesh faces, generating two weakly labeled datasets. We then introduce an instance-specific label propagation method that leverages geometric and topological cues to generate dense pseudo-labels from sparse annotations. Finally, we propose a robust label learning strategy that progressively exploits increasingly reliable samples and introduces a noise-suppression loss to improve pseudo-label quality during self-training. Extensive experiments on two widely used benchmarks, i.e. , COSEG and Human Body dataset, demonstrate that our method achieves performance comparable to state-of-the-art fully supervised approaches with 2% ∼ 5% annotation cost.

Segment anything model (SAM), a vision foundation network trained on a massive segmentation corpus, exhibits a superior boundary localization capability for nature images. This work aims to leverage such strengths to develop a deep unsupervised edge detection (UED) framework for alleviating the high reliance on dense labeling. However, applying vanilla SAM to edge detection fails to identify the salient edge cues but only the semantic boundary. This article introduces a lightweight adapter-tuning scheme to learn detailed edge information for filling the gap between boundary and edge, enabling a well-fitting even with limited training data. Moreover, considering the low-quality pseudo labels used in our UED framework, we propose two training strategies, adaptive progressive learning and gradient-guided pseudo label updating, to alleviate the impact of noisy labels from traditional UED methods. Extensive experiments demonstrate that our method achieves comparable results to state-of-the-art fully supervised edge detectors.

Unlabeled data is often used to improve the generalization ability of one segmentation model. However, it tends to neglect the inherent difficulty of unlabeled samples, and then produces inaccurate pseudo masks in some unseen scenes, resulting in severe confirmation bias and potential performance degradation. These motivate two unexplored questions for new-coming data: (1) How many images do we need to annotate; and (2) how to annotate them? In this paper, two kinds of shadow detectors (i.e., SDTR and SDTR+) based on the Transformer and self-training scheme are successively proposed. The main difference between them is whether weak annotations are required for partial unlabeled data. Specifically, in SDTR, we first introduce an image-level sample selection scheme to separate the unlabeled data into reliable and unreliable samples from the holistic prediction-level stability. Then, we perform selective retraining to exploit the unlabeled images progressively in a curriculum learning manner. While in SDTR+, we further provide various weak labels (i.e., point, box and scribble) for the rest unreliable samples and design corresponding loss functions. By doing this, it can achieve a better trade-off between performance improvement and annotation cost. Experimental results on public benchmarks (i.e., SBU, UCF and ISTD) show that both SDTR and SDTR+ can be favorable against state-of-the-art methods.

Compared to models pre-trained on ImageNet, the segment anything model (SAM) has been trained on a massive segmentation corpus, excelling in both generalization ability and boundary localization. However, these strengths are still insufficient to enhance shadow detection without additional training, and it raises the question: do we still need precise manual annotations to fine-tune SAM for high detection accuracy? This paper proposes an annotation-free framework for deep unsupervised shadow detection (USD) by leveraging SAM’s capabilities. The key lies in how to exploit the abilities acquired from a large-scale corpus and utilize them to improve downstream tasks. Instead of directly fine-tuning SAM, we propose a prompt-like tuning method to inject task-specific cues into SAM in a light-weight manner, namely ShadowSAM. This adaptation manner can ensure a good fitting when training data is limited. Moreover, considering that the pseudo labels used in our framework are generated by traditional USD approaches and may contain severe label noises, we propose an illumination and texture-guided updating strategy to selectively boost the quality of pseudo masks. To further improve the model’s robustness, we design a mask diversity index to establish easy-to-hard subsets for incremental curriculum learning. Extensive experiments on benchmark datasets (i.e., SBU, UCF, ISTD, and CUHK-Shadow) demonstrate that our unsupervised solution can achieve comparable performance to state-of-the-art (SOTA) fully supervised methods. Our code is available at this repository.