Zheng Luo, T Pranav Kutralingam, Ogochukwu N. Okoani, Wanpeng Xu, Hua Wei, Xiyang Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Unfortunately the funding information was incorrect. It should be the Key-Area Research and Development Program of Guangdong Province (No. 2019B010150001).

Telemedicine for 3D images on mobile devices presents promising development opportunities. Being constrained by computing power and storage capacity on mobile devices, the processing performance of 3D medical images is insufficient for more demanding tasks. Using virtual mobile infrastructure technology to utilize cloud resources is a common solution. But it encounters the challenge of poor performance in data transmission, image rendering and image coding. This paper presents a GPU and VPU enabled Open Virtual Mobile Infrastructure (OpenVMI) for 3D image rendering to solve the challenge. It makes two improvements. First, a bespoke GPU driver is developed in the Android Docker, optimizing the transmission workflow for data transmission and image rendering. Second, a Video Process Unit (VPU) is added to the hardware layer to code rendered results in H.264 format, replacing CPU coding which consumes a large amount of CPU resources. By adopting the OpenVMI, the telemedicine training system proposed in this paper presents an easy-to-set up, cheap and low latency solution that is particularly helpful for telemedicine training in remote and underdeveloped areas. Performance experiments suggest that the OpenVMI delivers better performance than existing state-of-the-art systems, even in mobile devices with weaker hardware capabilities. Concurrency experiment suggests that a single host server can support up to 24 concurrent training sessions, which makes the OpenVMI very helpful for telemedicine training that demands high concurrency. The OpenVMI-based solution proposed in this paper is not restricted to the use of telemedicine training, but also suitable for other application areas such as Virtual Reality and Augmented Reality in mobile environments.

Traffic simulation is an important computational technique that models the behavior and interactions of vehicles, pedestrians, and infrastructure in a transportation system. Calibration, which involves adjusting simulation parameters to match real-world data, is a key challenge in traffic simulation. Traffic simulators involve multiple models with hybrid hyperparameters, which could be either categorical or continuous. In this paper, we present CHy2, an approach that generates a set of hyperparameters for simulator calibration using generative adversarial imitation learning. CHy2 learns to mimic expert behavior models by rewarding hyperparameters that deceive a discriminator trained to classify policy-generated and expert trajectories. Specifically, we propose a hybrid architecture of actor-critic algorithms to handle the hybrid choices between hyperparameters. Experimental results show that CHy2 outperforms previous methods in calibrating traffic simulators.

Aiming for overcoming the constraints in graphic rendering of mobile devices used in telemedicine training for 3D images, this paper presents a GPU-enabled mobile telemedicine training system for graphic rendering. Two improvements are made to seamlessly display interactive 3D images of human organs, bones and blood vessels with realtime rendering. Firstly, instead of multiple stages of instruction translation between OpenGL ES and OpenGL, a bespoke GPU driver is added in Virtual Android to invoke GPU resources directly. Secondly, a Video Process Unit (VPU) is added to the hardware layer replacing CPU to code rendered results in H.264 format, reducing CPU load significantly. Test results suggest that the system can deliver consistent performance even in mobile devices of weak capability and a single server can support up to 24 concurrent virtual Android operating systems, each of which connects to 5 clients. The framework proposed by this paper is not only suitable for telemedicine training, but also for other application areas such as Virtual Reality and Augmented Reality in mobile environment.

Voxel-based structures in 3D detection have achieved rapid advancement due to their superior capability for feature extraction. However, the accuracy is usually low because the point cloud is divided into a grid. In order to overcome the above problems and improve detection accuracy, we propose a flexible two-stage 3D object detection architecture, which adopts two branches to refine generated proposals, aggregating voxel features and raw point features simultaneously. We also design a new gating mechanism to achieve fusion features from different levels. In addition, we propose a novel feature aggregation module to reduce the semantic gap between the features of the two types. First, a transformer based on raw points is employed as an encoder to aggregate the contextual information. Then, the point-based channel-wise self-attention mechanism serves as a decoder to aggregate the global features. Experiment results on the KITTI 3D dataset and Waymo Open datest demonstrate that our approach outperforms the state-of-the-art methods and exhibits excellent scalability.

Self-supervised depth estimation approaches present excellent results that are comparable to those of the fully supervised approaches, by employing view synthesis between the target and reference images in the training data. ResNet, which serves as a backbone network, has some structural deficiencies when applied to downstream fields, because its original purpose was to cope with classification problems. The low-texture area also deteriorates the performance. To address these problems, we propose a set of improvements that lead to superior predictions. First, we boost the information flow in the network and improve the ability to learn spatial structures by improving the network structures. Second, we use a binary mask to remove the pixels in low-texture areas between the target and reference images to more accurately reconstruct the image. Finally, we input the target and reference images randomly to expand the dataset and pre-train it on ImageNet, so that the model obtains a favorable general feature representation. We demonstrate state-of-the-art performance on an Eigen split of the KITTI driving dataset using stereo pairs.