Real-time Occlusion Rendering

This demo shows how to use ARCore's Depth API for occlusion rendering in real-time. A 3D teapot mesh is rendered such that it appears behind real-world geometry using single-frame depth estimation from the ARCore SDK.

• Depth is not integrated across frames — this is a minimal, focused implementation
• The purpose is to demonstrate clean occlusion without relying on plane detection or complex scene understanding

Demo

Video on project page

Lessons Learned

• ARCore depth and pose access require an active GL context. This means the usual ImageAnalysis thread doesn't work — depth access must happen in a coroutine tied to the GL rendering thread.
• ARCore camera feed is exclusive — cannot use CameraX alongside ARCore.
• A recomposing Jetpack Compose setup will recreate Session unless it is remembered — which breaks texture attachments and depth acquisition.
• Correct occlusion rendering requires exact matching of glViewport and display geometry, especially across screen rotations. Otherwise, the mesh appears squished or misaligned.
• Visual alignment between the depth map and RGB frame was verified manually. Frame-to-depth transformation was skipped in favor of a fixed rotation for simplicity.

---

Keywords: ARCore, Android, Kotlin, OpenGL ES, Depth Map, Pose Estimation, Occlusion, 3D Rendering, Augmented Reality, Jetpack Compose, Real-Time Graphics, Teapot Mesh, AR Visualization, Mobile AR