Academic publications in Artificial Intelligence and Computer Science.
A novel multi-modal training scheme that leverages LiDAR geometric details to enhance image-guided depth completion for autonomous driving. The proposed BEV@DC model achieves state-of-the-art performance, ranking Top-1 on the challenging KITTI depth completion benchmark.
This thesis addresses two major challenges in RL for 3D environments: high memory consumption and POMDP complexity. We propose novel SS-only and RGB+SS input representations using Semantic Segmentation, achieving up to 98.6% memory reduction with run-length encoding applied to SS-only while significantly enhancing agent performance in ViZDoom deathmatches with RGB+SS.
This study addresses the research question of where to place examples (shots) in LLM prompts to improve performance. We systematically test different shooting combinations with system and user messages to determine the optimal placement for enhancing LLM performance.