Accurately distinguishing between action-related features and non-action-related features is crucial in spatio-temporal action detection tasks. Additionally, the calibration and fusion of information across different modalities remain challenging. This paper proposes a novel Motion Sense and Semantic Correction framework (MS-SC) to address these issues. The MS-SC framework achieves accurate detection by fusing features from images (spatial dimension) and videos (spatio-temporal dimension). A Motion Sense Module (MSM) is proposed to significantly increase the feature distance between action and non-action features in the semantic space, enhancing feature discriminability. Considering the complementary nature of information across different modalities, an efficient Semantic Correction Fusion Module (SFM) is introduced to facilitate interaction between features of distinct modalities and maximize their complementary information integration. To evaluate the performance of the MS-SC framework, extensive experiments were conducted on two challenging datasets, UCF101-24 and AVA. The results demonstrate the effectiveness of the MS-SC framework in handling spatio-temporal action detection tasks.

This work aims to investigate the problem of 3D modeling using single free-hand sketches, which is one of the most natural ways we humans express ideas. Although sketch-based 3D modeling can drastically make the 3D modeling process more accessible, the sparsity and ambiguity of sketches bring significant challenges for creating high-fidelity 3D models that reflect the creators’ ideas. In this work, we propose a view-and structural-aware deep learning approach, Deep3DSketch, which tackles the ambiguity and fully uses sparse information of sketches, emphasizing the structural information. Specifically, we introduced random pose sampling on both 3D shapes and 2D silhouettes, and an adversarial training scheme with an effective progressive discriminator to facilitate learning of the shape structures. Extensive experiments demonstrated the effectiveness of our approach, which outperforms existing methods – with state-of-the-art (SOTA) performance on both synthetic and real datasets.