Single-view 3D reconstruction in complex real-world scenes is challenging due to noise, object diversity, and limited dataset availability. To address these challenges, we propose MGP-KAD, a novel multimodal feature fusion framework that integrates RGB and geometric prior to enhance reconstruction accuracy. The geometric prior is generated by sampling and clustering ground-truth object data, producing class-level features that dynamically adjust during training to improve geometric understanding. Additionally, we introduce a hybrid decoder based on Kolmogorov-Arnold Networks (KAN) to overcome the limitations of traditional linear decoders in processing complex multimodal inputs. Extensive experiments on the Pix3D dataset demonstrate that MGP-KAD achieves state-of-the-art (SOTA) performance, significantly improving geometric integrity, smoothness, and detail preservation. Our work provides a robust and effective solution for advancing single-view 3D reconstruction in complex scenes.

Knowledge distillation is a popular technique for transferring the knowledge of a teacher model to a smaller and more efficient student model. However, previous work often used certain intermediate features as knowledge sources, which greatly reduced the generality of model distillation on heterogeneous models. To address this challenge, we propose a novel knowledge distillation method called Knowledge Calibration Distillation (KCD) for efficient transfer learning between teacher and student models. KCD incorporates three key innovations: category uncertainty calibration, adaptive cosine-similarity distillation, and dynamic temperature factors. The category uncertainty calibration mechanism prevents the student from learning from unreliable teacher predictions. The adaptive cosine-similarity distillation loss functions capture comprehensive knowledge, including category and batch information. Dynamic temperature factors optimize the distillation process by adapting to the similarity between teacher and student outputs. Extensive experiments on the CIFAR-100 dataset demonstrate that KCD consistently outperforms existing knowledge distillation methods across various teacher-student model combinations, achieving state-of-the-art results in terms of accuracy1

This paper proposes a method to complete the missing 3D point cloud reconstructed from aerial multi-view images by using a deep learning method with self-attention. The advancement of drone technology has made it easier to acquire aerial multi-view images. While it is possible to generate 3D point clouds of the terrain by applying 3D photogrammetric techniques to these images, when capturing multi-view aerial images with a drone, high-altitude vertical shooting is often necessary for privacy protection. For example, some portions of the generated 3D point clouds are lost due to shadowed areas caused by roofs and eaves. To address this issue, this research proposes a method to complete the missing 3D point cloud by using a deep learning. In order to obtain accurate and sufficient amount of training data, 3D CG building models are used for generating sets of missing 3D point cloud data and their corresponding Ground Truth. In the experiment, we applied our method to a 3D point cloud generated from actual captured aerial multi-view images and confirmed that the point cloud with a reasonable shape for the missing parts are successfully completed.

One of the major challenges in virtual reality (VR) is to create a perception of a large virtual space within a limited physical space. Here, we explore the effect of haptic-based navigation by Hanger Reflex (HR) on the perception in redirected walking (RDW) with visual manipulation. Seven individuals walked along a straight path in VR while, unbeknown to them, the visual scene was rotated with a curvature gain of π/36, forcing them to walk in a circular path in real space. HR rotation (in the left, right, and neutral direction) was induced by a wearable haptic device during each of the walking trials, and they reported their perceived walking direction and effort to walk along the path on visual analog scale. Experiment results showed that HR can influence the perception in RDW, but the effects may be complex and therefore require further investigation.

This paper proposes a calibration method for a large-scale floor projection system at a school gym. In our system, the projectors are installed on the ceiling of a school gym, and the contents are projected onto the floor. The projection results suffer from both projective distortion and lens distortion. It is hard to use chessboard-based calibration methods or self-calibration methods in our system due to the restrictive deploy environment and practical limitations. Our method is basing on the projector-camera system and "straight lines have to be straight". The experiments show that our method fulfills the on-site requirements of the floor projection system at school gym for learning aid usages.

We propose a calibration method for deploying a large-scale projection-based floor display system. In our system, multiple projectors are installed on the ceiling of a large indoor space like a gymnasium to achieve a large projection area on the floor. The projection results suffer from both perspective distortion and lens distortion. In this paper, we use projector-camera systems, in which a camera is mounted on each projector, with the “straight lines have to be straight” methodology, to calibrate our projection system. Different from conventional approaches, our method does not use any calibration board and makes no requirement on the overlapping among the projections and the cameras' fields of view.