It is often the case that data are with multiple views in real-world applications. Fully exploring the information of each view is significant for making data more representative. However, real data tend to suffer from arbitrary view-missing due to various failures in data collection and pre-processing. Besides, as obtaining large-scale labeled data is laborious and expensive, the collected training samples in the target task may be scarce as well. The co-existence of these two problems makes it more challenging to achieve the pattern classification task. Currently, to our best knowledge, few appropriate methods can well-handle these two problems simultaneously. To draw more attention from the community to this challenge, we present a new task in this paper called Few-shot Partial Multi-view Learning (FPML), aiming to alleviate the view-missing effects in the low-data regime. The challenges of this task are twofold: (1) under the interference of the missing views, it is difficult to overcome the negative impact brought by data scarcity; (2) the limited number of data exacerbates information scarcity, thereby making it harder to address the view-missing problem. In this paper, we propose a novel method called Unified Gaussian Dense-anchoring (UGD) for this task. We propose to learn the unified dense Gaussian anchors for each training sample. Therefore, the incomplete multi-view data can be densely anchored into a unified representation space, where data scarcity and view missing are both relieved by the unified dense anchor representations. In particular, we also extend our FPML to the multimodal and the cross-domain scenario, and validate our UGD method on them. The extensive experiments firmly demonstrate the effectiveness of our approach.