Development Blog
AABB-based Culling Optimizations in Real-Time Rendering Engines
Current real-time rendering engines often struggle with visibility bottlenecks that limit frame rates and scalability. We propose an optimized framework that combines hierarchical culling, GPU-driven occlusion queries, and caching strategies to deliver up to a tenfold performance gain. These results highlight new opportunities for extending visibility algorithms and invite deeper exploration in the full paper.
Region Aware Navigaton using Recast and Detour
This paper introduces a regionization and path-smoothing pipeline for navigation meshes to improve stability and runtime queries. Here, regions are generated from tiled navigation meshes generated using Recast builds by sampling distance-to-wall, discretizing cells and labeling with a descending level sweep and BFS. We explore path smoothing by using portal funnel algorithm, with optional rules to penalize or restrict movement across region boundaries, with the expected result producing a taut route with minimal corners and LOS optimized segments.
Generating 3D Bounding Boxes with 2D Object Detection
Current object detection algorithms are capable of generating axis-aligned bounding boxes for 2D images. However, such a technique does not currently exist for 3D objects. Although 2D object detection can be performed on a scene viewed from a specific point, it cannot generate a 3D axis-aligned bounding box containing the specified objects. With the growth of artificial intelligence, performing detection on 3D objects within a scene offers countless potential uses. We propose a technique to generate 3D axis-aligned bounding boxes for 3D objects stored as a triangle mesh by performing 2D object detection on the scene viewed from a variety of angles and then assigning triangles to the detected categories using ray-tracing on a BSP (binary space partitioning) tree constructed from the scene.
