Recently, the IEEE International Conference on Computer Vision (ICCV2021), the top conference in the field of artificial intelligence, was officially held, and the results of this year's ICCV competition were officially announced. The International Joint Laboratory for Pattern Recognition and Computational Intelligence (IJLPRCI), led by Professor Wu Xiaojun of the School of Artificial Intelligence and Computer Science at Jiangnan University and Professor Josef Kittler of the Department of Electrical Engineering at the University of Surrey, UK, and Professor Jiwen Lu of the Department of Automation of Tsinghua University, won the Multi-Modal Video Reasoning and Analyzing Competition (MMVRAC) 2 winners and 3rd runner-up in the 2nd Anti-Drone Anti-UAV Tracking Competition.
<< Swipe to see the next image >>
In the bone-based behavior recognition, the team made full use of the differences between features, scales and different modeling methods to complement each other's advantages in view of the difficulties such as large changes in visual angles, high similarity of some sample categories, and limited data scale, and enhanced the classification performance of the model through the integration of classification features.
In the behavior recognition based on fisheye video, the team used a variety of data processing methods to reduce lens distortion, focus on moving subjects, and reduce the impact of drone lens shake and complex lighting scenes. In the model design, multi-scale sampling methods are used and corresponding modeling is adopted. During the testing phase, the team used multi-perspective test enhancements, decision-level fusion of multi-variety sampling models, etc. to improve the final classification accuracy.
In the anti-drone tracking problem, the team proposed a long-term tracking solution for the difficulties of fast target movement, rapid movement of sampling equipment, repeated disappearance and reproduction of targets, and integrated tracking algorithms based on Transformer and twin networks. In addition, for the problem of disappearing and reproducing targets, the target re-detection method of global search is used.
The skeleton-based motion recognition and fisheye video-based action recognition competitions are based on the latest and challenging drone perspective video understanding dataset, UAV-Human, focusing on understanding and reasoning human behavior from the drone perspective, including 67,428 video samples, 6 different modes, 4 human behavior understanding tasks and 119 video themes.
In the face of large and complex data samples, Professor Wu Xiaojun's team needed to upgrade its existing underlying computing power in order to obtain more algorithm optimization time and improve model performance to achieve test training time compression.
However, with the current high-speed upgrade of computing power, GPU power consumption continues to rise, it is difficult for traditional air-cooled workstations to achieve efficient heat dissipation of multi-card workstations, and there is a risk of frequency reduction, which can easily lead to overall performance degradation. Moreover, in order to carry the multi-card heat dissipation, the multi-fan stacking with larger air volume and higher wind pressure has also promoted the noise of the Doka workstation to become higher and higher (the full load can reach more than 65dB), and it is no longer possible to directly place in the office environment.
In order to help Professor Wu Xiaojun's team solve the current workstation frequency reduction and noise problems, Chaoyi Jilun provided it with a full liquid cooling workstation solution - ServMAX® TL40-X2. The whole machine adopts CPU + GPU full liquid cooling design, the noise is less than 55dB, and the GPU core runs stably below 72 °C under room temperature conditions, which effectively avoids the risk of frequency reduction.
<h1 class="pgc-h-arrow-right" data-track="14" > dual processor, quad flagship GPU</h1>
The whole machine can reach 56 cores and 112 threads
4-channel GPU, overall video memory up to 96GB
Support NVLink, bandwidth up to 112GB/s
Parallel GPU, support flexible disassembly
<h1 class="pgc-h-arrow-right" data-track="41" > full liquid cold heat dissipation, extreme noise reduction</h1>
The full load of the whole machine noise is within 55dB
Circulating air duct design, heat dissipation without dead angles
The fan speed is adjustable, seeking a balance between efficiency and silence
Closed circulating waterway, no risk of leakage
<h1 class="pgc-h-arrow-right" data-track="42" > flexible configuration and high scalability</h1>
Supports 10 x 3.5"/2.5" + 2 x 2.5" SATA/SAS internal hard drives
Supports NVMe M.2 (2280/22110) SSDs
Comes with multiple USB 3.0 and USB 2.0 ports
Dual on-board 10Gbps RJ45 Ethernet ports and IPMI management ports
<h1 class="pgc-h-arrow-right" data-track="43" > intelligent LCD panel for real-time status monitoring</h1>
Intuitively grasp key temperatures
Real-time monitoring of several important parameters of the cooling system
After repeated testing and comparison, the ServMAX® TL40-X2 finally chose gpu parallel heat dissipation. After the water body is cooled, at the same time, it will flow through all the GPUs in parallel, and after the heat is taken away as a whole, it will be returned to the cold drain for heat dissipation, which ensures that all GPUs are in a low temperature state. At the same time, it adopts a reinforced connector and a water stop valve design to ensure that any GPU can be flexibly installed or removed. Secondly, the water is transmitted through the hose, which avoids the risk of fracture and long-term leakage of most hard water pipes on the market without affecting the flow rate and heat dissipation effect.