The 3D bin packing problem has aroused enthusiastic research interest in recent years due to its wide range of real-world applications, such as logistics and warehousing. The packing sequence and placement pose (position and orientation) are the primary optimization objectives in bin packing, as they dramatically impact packing results and transportation costs. Existing methods either focus exclusively on sequential or placement decisions, or navigate a vast combined action space of both, generally struggling to reach global optimum. To bridge this research gap, we propose a novel approach that jointly optimizes both objectives. First, we introduce a packing configuration tree to represent the dynamic packing process. Second, we develop a hierarchical reinforcement learning framework which decomposes the problem into two tractable subtasks: a high-level manager network for sequence generation and a low-level worker network for placement determination. Third, we formulate a stepwise reward and train the framework using a Dueling Deep Q-Network. Finally, extensive experiments demonstrate the superiority of the proposed method, which achieves over 10% higher container space utilization compared to the advanced methods while maintaining robust generalization across varied container and box dimensions. Furthermore, real-world robotic packing experiments validate the practical applicability of the proposed method in industrial scenarios.

Multi-modal entity alignment aims to identify equivalent entities across diverse knowledge graphs by leveraging multiple modalities of entity information. This process is crucial for the fusion of multi-modal knowledge graphs. While current research primarily investigates how to utilize side information from entity visuals, relations, and attributes, it often overlooks the significant role of entity-type information. Furthermore, multi-modal data embedding encounters noise that negatively impacts the performance of the entity alignment task. To address these gaps, this paper introduces MTCEA, a multi-modal entity alignment method guided by entity-type information. The proposed method captures the constraints associated with entities based on the entity-type information obtained from knowledge graph ontology; then, it utilizes two embedding strategies for type constraints to enhance the model’s performance in knowledge representation. This allows effective modal fusion that integrates more fine-grained semantic constraints related to types, which improves the alignment accuracy across various cross-lingual knowledge graphs. MTCEA is validated on three subsets of DBP15K. Experimental results demonstrate that our model achieves good results overall on the Hits@1, Hits@10, and MRR metrics. In an experimental setting without using entity name, MTCEA outperforms state-of-the-art baselines.

Due to the inherent characteristics of the visualization process, most of the problems in this field have strong ties with human cognition and perception. This makes the human brain and sensory system the only truly appropriate evaluation platform for evaluating and fine-tuning a new visualization method or paradigm. However, getting humans to volunteer for these purposes has always been a significant obstacle, and thus this phase of the development process has traditionally formed a bottleneck, slowing down progress in visualization research. We propose to take advantage of the newly emerging field of Human Computation (HC) to overcome these challenges. HC promotes the idea that rather than considering humans as users of the computational system, they can be made part of a hybrid computational loop consisting of traditional computation resources and the human brain and sensory system. This approach is particularly successful in cases where part of the computational problem is considered intractable using known computer algorithms but is trivial to common sense human knowledge. In this paper, we focus on HC from the perspective of solving visualization problems and also outline a framework by which humans can be easily seduced to volunteer their HC resources. We introduce a purpose-driven game titled "Disguise" which serves as a prototypical example for how the evaluation of visualization algorithms can be mapped into a fun and addicting activity, allowing this task to be accomplished in an extensive yet cost effective way. Finally, we sketch out a framework that transcends from the pure evaluation of existing visualization methods to the design of a new one.

The unguided visual exploration of volumetric data can be both a challenging and a time-consuming undertaking. Identifying a set of favorable vantage points at which to start exploratory expeditions can greatly reduce this effort and can also ensure that no important structures are being missed. Recent research efforts have focused on entropy-based viewpoint selection criteria that depend on scalar values describing the structures of interest. In contrast, we propose a viewpoint suggestion pipeline that is based on feature-clustering in high-dimensional space. We use gradient/normal variation as a metric to identify interesting local events and then cluster these via k-means to detect important salient composite features. Next, we compute the maximum possible exposure of these composite feature for different viewpoints and calculate a 2D entropy map parameterized in longitude and latitude to point out promising view orientations. Superimposed onto an interactive track-ball interface, users can then directly use this entropy map to quickly navigate to potentially interesting viewpoints where visibility-based transfer functions can be employed to generate volume renderings that minimize occlusions. To give full exploration freedom to the user, the entropy map is updated on the fly whenever a view has been selected, pointing to new and promising but so far unseen view directions. Alternatively, our system can also use a set-cover optimization algorithm to provide a minimal set of views needed to observe all features. The views so generated could then be saved into a list for further inspection or into a gallery for a summary presentation.

Practical volume visualization pipelines are never without compromises and errors. A delicate and often-studied component is the interpolation of off-grid samples, where aliasing can lead to misleading artifacts and blurring, potentially hiding fine details of critical importance. The verifiable visualization framework we describe aims to account for these errors directly in the volume generation stage, and we specifically target volumetric data obtained via computed tomography (CT) reconstruction. In this case the raw data are the X-ray projections obtained from the scanner and the volume data generation process is the CT algorithm. Our framework informs the CT reconstruction process of the specific filter intended for interpolation in the subsequent visualization process, and this in turn ensures an accurate interpolation there at a set tolerance. Here, we focus on fast trilinear interpolation in conjunction with an octree-type mixed resolution volume representation without T-junctions. Efficient rendering is achieved by a space-efficient and locality-optimized representation, which can straightforwardly exploit fast fixed-function pipelines on GPUs.