Reconstructing neural radiance fields from limited or sparse views has given very promising potential for this field of research. Previous methods usually constrain the reconstruction process with additional priors, e.g. semantic-based or patch-based regularization. Nevertheless, such regularization is given to the synthesis of unseen views, which may not effectively assist the field of learning, in particular when the training views are sparse. Instead, we propose a feature Field Fusion (FFusion) NeRF in this paper that can learn structure and more details from features extracted from pre-trained neural networks for the sparse training views, and use as extra guide for the training of the RGB field. With such extra feature guides, FFusion predicts more accurate color and density when synthesizing novel views. Experimental results have shown that FFusion can effectively improve the quality of the synthesized novel views with only limited or sparse inputs. • A novel NeRF method using spatial feature maps to improve sparse view synthesis. • Concurrently learning both feature & color fields with pixel-level loss supervisions. • A fusion scheme regularizing color prediction with learned features as condition guide. • Extensive tests validate the performance and model-agnostic advantages of the method.

Deep learning (DL)-based human activity recognition (HAR) has attracted considerable attention owing to its vast potential across various applications. Currently, HAR still faces two challenges. For one thing, existing methods neglect the spatial distribution information embedded in HAR signals and lack the ability to comprehensively model the spatial-temporal (ST) dependencies within HAR data, restricting them from effectively decoding human activity. For another thing, previous models generate feature representations for a sliding window of the sequence solely based on this window itself, without cross-window interaction learning, posing challenges to classifiers, such as perceptual aliasing or feature inconsistency issues. For that, we propose a novel cross-window and cross-modality correlation-aware framework, namely LCwmcaR, which is a dual-branch network that simultaneously models temporal- and spatial-level information using Mamba and convolutional neural network (CNN), respectively. Additionally, a learnable temporal 2-dimensionalization (LT2D) strategy is designed to encode low-level temporal patterns into high-level learnable image-liked 2-D space representations that integrate both local and global ST dependencies. Moreover, a cross-window correlation-aware feature representation generation (CrwcaFRGen) module, which correlates multiple windows representations within a batch at the attention level, is introduced to produce more robust features for the classifier. Experimental results on four public datasets demonstrate that the LCwmcaR outperforms state-of-the-art (SOTA) methods by a large margin.