Demos

Challenge: The high cost and time-consuming nature of annotating and training 3D LiDAR point cloud data.

Approach: We propose a pipeline with a lightweight while efficient architecture (LiM3D), with the Sparse Depthwise Separable Convolution (SDSC) module and Spatio-Temporal Redundant Frame Downsampling (ST-RFD) to reduce training parameters and enhance 3D semantic segmentation performance with minimal annotations.

Application: Beyond autonomous driving, our method can be applied to other fields requiring efficient semantic segmentation, such as robotics, urban planning, and augmented reality. Future research could explore extending this approach to different sensor modalities and more diverse environmental conditions.

Challenge: Existing 3D LiDAR segmentation methods predominantly rely on point coordinates and intensity, which, while effective under stable conditions, exhibit poor isometric invariance and struggle with varying point densities, leading to suboptimal semantic segmentation in complex outdoor environments.

Approach: We propose RAPiD features, a novel representation capturing localized geometry through a 4D distance metric that integrates spatial structure and material reflectivity. These features are embedded via a double-nested autoencoder with a class-aware objective, enabling efficient voxel-wise encoding. We further introduce RAPiD-Seg, an architecture leveraging channel-wise attention fusion and tailored variants to maximize segmentation performance.

Application: Our method is tailored for outdoor 3D scene understanding, particularly in autonomous driving scenarios, and demonstrates state-of-the-art mIoU performance on SemanticKITTI (76.1) and nuScenes (83.6) benchmarks.

Challenge: Achieving high-resolution, accurate perception performance with limited ground truth data remains a significant challenge in autonomous driving due to the expense and complexity of annotating 3D LiDAR point clouds.

Approach: We introduce DurLAR, a high-fidelity 128-channel 3D LiDAR dataset with panoramic ambient and reflectivity imagery. With a novel proposed joint supervised/self-supervised loss formulation, we achieve SoTA results for monocular depth estimation.

Application: This work can benefit fields such as robotics, urban planning, and augmented reality, where precise depth and high-resolution LiDAR data is crucial. Future research could explore integrating additional sensor data and enhancing performance in diverse environmental conditions.

Challenge: AI-based chatbots have become increasingly popular in various domains. Current chatbots primarily target younger users, leaving a gap in understanding and catering to the music preferences and needs of older adults.

Approach: We conduct a survey with 20 older adults to gather insights into their music preferences and specific requirements for a music chatbot. Based on the findings from the survey, we develop a prototype music chatbot tailored to these identified needs. We then test the prototype with five older adults to gather feedback and assess its effectiveness and usability.

Application: It holds significant potential for enhancing various aspects of older adults' lives, including healthcare support, mental health companionship, and social connectivity. Future work should focus on improving accessibility, personalization, and emotional intelligence to create more inclusive and empathetic interactions for this demographic.

Challenge: Writing Bash files from scratch for SLURM is complex and time-consuming, especially for beginners. It is error-prone, requires frequent modifications, and finding appropriate configurations is tedious. Updates are inconvenient, feedback is delayed, and heavy documentation dependence adds to the challenge.

Approach: Our graphical interface for SLURM simplifies configuration with a user-friendly design, saving time through predefined settings and minimizing errors with dropdown menus. It supports real-time interaction, reusability of past configurations, and online debugging, enhancing accessibility across devices.

Application: The graphical HPC helper streamlines computational tasks in research and industry, simplifies learning for new users, and enhances rapid prototyping with real-time interaction and ease of use.