AI Cluster - Team of Prof. Xiaoguang HAN (Provided by Zhongjin Luo)
SimpModeling: Sketching Implicit Field to Guide Mesh Modeling for 3D Animalmorphic Head Design
Head shapes play an important role in 3D character design. In this work, we propose SimpModeling, a novel sketch-based system for helping users, especially amateur users, easily model 3D animalmorphic heads - a prevalent kind of heads in character design. Although sketching provides an easy way to depict desired shapes, it is challenging to infer dense
geometric information from sparse line drawings. Recently, deepnet-based approaches have been taken to address this challenge and try to produce rich geometric details from very few strokes. However, while such methods reduce users' workload, they would cause less controllability of target shapes. This is mainly due to the uncertainty of the neural prediction. Our
system tackles this issue and provides good controllability from three aspects: 1) we separate coarse shape design and geometric detail specification into two stages and respectively provide different sketching means; 2) in coarse shape designing, sketches are used for both shape inference and geometric constraints to determine global geometry, and in geometric detail crafting, sketches are used for carving surface details; 3) in both stages, we use the advanced implicit-based shape inference methods, which have strong ability to handle the domain gap between freehand sketches and synthetic ones used for training. Experimental results confirm the effectiveness of our method and the usability of our interactive system. We also contribute to
a dataset of high-quality 3D animal heads, which are manually created by artists.
The major contributions of this work are summarized as follows:
- We design an easy-to-use and controllable sketch modeling interface for animalmorphic head modeling. In particular, it takes around ten minutes for novice users to create a desired animalmorphic head model using only a small number of 3D strokes.
- We propose a coarse-to-fine shape inference method, which seamlessly integrates explicit and implicit shape representations to guarantee reconstruction quality and efficiency.
- We contribute to the largest animalmorphic head dataset that consists of 1,955 high-quality animalmorphic head models. Each model in the proposed dataset is carved manually by artists and carefully annotated with 3D contour annotations. We will make the dataset together with the sketch modeling system publicly available to the research community.
CE Cluster - Team of Prof. Shuk-Yin TONG (Provided by Fangrun Jin)
Hydrogen energy is considered as one of the ideal green energy sources due to its superior features such as high energy density and low pollution. The solar hydrogen generation via photocatalytic water splitting has attracted tremendous attention since it has great potential for low-cost and clean hydrogen production. Recently, the water dissociation on TiO2 has been intensively studied and plenty of impressive works have been done. However, for the surface reaction mechanism of this process, there still remain some bottlenecks to solve. In this research we apply first principal calculations using density functional theory (DFT) with HSE 06 functional on the high-performance computational platform of Cluster Engine to comprehensively study the change of adsorption energy with different coverages of hydrogen on rutile TiO2 (110) surface and the localized sites of electrons. Additionally, other important information including energy band and DOS are also discussed. Therefore, the theory study can provide guidance and support for the possible future hydrogen generation application.




CE Cluster - Team of Prof. Ying-Chih CHIANG (Provided by Muhui Ye)
Alzheimer's disease is associated with the aggregation of β-amyloid peptides (Aβ40, Fig. 1) in the brain. Our collaborators have designed a short cyclic peptide to inhibit this aggregation process.

Fig. 1
They then employed cryogenic electron microscopy (cryo-EM) to determine how the cyclic peptide can bind to Aβ40. Fig. 2 shows the image taken by the cryo-EM, where Aβ40 can be modeled in the middle (cyan), but the exact location of the cyclic peptide is uncertain: It could be at the site labeled by the asterisks (***) or at the site labeled by the dollar sign ($).

Fig. 2
To associate the cryo-EM image with the cyclic peptide, we conducted molecular dynamics (MD) simulations to investigate how the cyclic peptide binds to the Aβ40. MD simulates a microscopic system based on Newton’s laws and provides insights into a chemical or biological process with atomistic details. The calculations were carried out on the high-performance cluster at The Chinese University of Hong Kong (Shenzhen).
As shown in Video 1, the cyclic peptide unfolds a little and binds to Aβ40 in parallel to the β-amyloid peptide. Superimposing the simulation result and the cryo-EM image, we see that the cyclic peptide overlaps well with the density labeled by the asterisks (***), demonstrating that this is where the cyclic peptide binds. To sum up, this study helps to reveal the mechanism of how an anti-Alzheimer drug functions.

Fig. 3
AI Cluster - Team of Prof. Tin Lun LAM (Provided by Jingtao Tang)
For large-scale tasks, a multi-robot system can effectively improve efficiency, especially for heterogeneous multi-robot system (HMRS), the system can benefit more from utilizing different capabilities, mobility and functionality of each robot in the HMRS.
In this research, we introduce a worker-station HMRS consisting of multiple workers with limited energy or consumable capacity for actual work and one station with enough resources for supply replenishment to solve the multi-robot coverage path planning (mCPP) problem.
We design an efficient modeling approach of the mCPP problem for worker-station HMRS, and propose a novel end-to-end decentralized online planning approach based on Deep Reinforcement Learning (DRL), which solves coverage planning for workers as well as rendezvous planning for station, with the ability of collision avoidance with dynamic interferers in the environment.
With the help of the AirStation with massive training capability, our DRL based method successfully generates cooperative coordination behaviors towards the coverage task for each robot in the HMRS, leading to a satisfying and competitive performance.


AI Cluster - Team of Prof. Xiaoguang HAN (Provided by Kenkun Liu)
This video shows the effect of a newest human pose estimation algorithm. In this video, the upper part is the input video stream of the algorithm, and the lower part is the output result of the algorithm. It can be seen that this algorithm completely captures the actions of all people appearing in the input video and presents it in the form of a human skeleton. The algorithm does not need to know in advance how many people are in the video and where each person is. Its only input is the video stream, and then things are done automatically. This algorithm can be applied to many everyday scenarios, such as analyzing whether the actions of fitness enthusiasts are standard, automatically detecting whether the actions of players in sports games are foul, and automatically detecting whether the elderly living alone fall down, and so on. In the process of training the algorithm, I used 4 GPUs of the GeForce RTX 2080 Ti provided by AI Station as the computing resources for parameter training, and then used a single such GPU to achieve an average computing speed of 0.3 second per frame when making predictions.