AR environment occlusion effects implementation

Implementing AR Environment Occlusion Effects

Environment occlusion is when real objects properly occlude virtual ones. Placed AR sofa against wall, walk behind it — sofa partially hides behind your silhouette. Without occlusion AR object always draws on top — and presence illusion breaks immediately. Technically one of most demanding effects because it needs real-world depth information in real-time.

Two Fundamentally Different Paths to Occlusion

LiDAR path (iPhone 12 Pro+, iPad Pro): ARWorldTrackingConfiguration.sceneReconstruction = .mesh builds dense environment mesh. This mesh writes values to depth buffer. AR object behind mesh is occluded. Accuracy to 1–2 cm at distance up to 5 meters. Real-time no delay.

ML path (any ARKit/ARCore device): neural network predicts depth map from RGB image (monocular depth estimation). ARKit ARDepthMap / ARCore Depth API (DepthPoint). Accuracy worse (5–15 cm), artifacts at object edges, slight 1–3 frame delay.

For production: if audience is iPhone Pro / iPad Pro, use LiDAR. For mass app — ML depth with LiDAR acceleration where available.

RealityKit: Enable Occlusion One Line

On devices with LiDAR and iOS 15+:

arView.environment.sceneUnderstanding.options = [
    .occlusion,        // real objects occlude AR
    .collision,        // AR objects collide with real geometry
    .physics,          // physics relative to real surfaces
    .receivesLighting  // AR objects lit like real
]

This is literally all needed for basic occlusion in RealityKit. Under the hood — ARKit builds environment mesh, RealityKit uses it as occluder.

Without LiDAR on same API: ARView uses ARKit person segmentation (A12+) — only human silhouette occludes AR objects. Table, chair, wall — don't occlude. Limited, but better than nothing.

Person Occlusion: Separate Task

ARWorldTrackingConfiguration.frameSemantics = .personSegmentation — ARKit segments person on frame and creates stencil mask. AR objects "behind" person occluded by their silhouette. Works on A12+ without LiDAR.

Mask quality — small halo 2–5 pixels around body edge. On dark background noticeable, on natural — acceptable.

For custom rendering: ARFrame.segmentationBuffer — CVPixelBuffer with pixel classes (ARFrame.SegmentationClass.person). Used in Metal shaders for custom compositing.

ARCore Depth API (Android)

On devices with ToF sensor (Samsung Galaxy S21 Ultra, LG V60) — hardware depth. On rest — ML depth from ARCore Depth API (Raw Depth Image):

val frame = session.update()
if (session.isDepthModeSupported(Config.DepthMode.AUTOMATIC)) {
    val depthImage = frame.acquireDepthImage16Bits()
    // depth in mm, 16-bit unsigned short
    // ARCore recommends use via OpenGL texture
}

ARCore provides OcclusionShader via sample — OpenGL fragment shader discarding AR object fragments deeper than real geometry. Foundation for occlusion in Sceneform/custom renderer.

Custom Occlusion Shader (Metal / OpenGL ES)

For non-standard requirements (need blurred edge on occlusion boundary, artistic style) — custom Metal shader:

fragment float4 occlusionFragment(
    VertexOut in [[stage_in]],
    texture2d<float> arDepthMap [[texture(0)]],
    texture2d<float> vrObject [[texture(1)]]
) {
    float2 uv = in.texCoord;
    float realDepth = arDepthMap.sample(s, uv).r; // depth of real scene
    float vrDepth = in.depth; // depth of AR object

    if (vrDepth > realDepth + 0.01) {
        discard_fragment(); // AR object behind real geometry
    }

    return vrObject.sample(s, uv);
}

Offset 0.01 (1 cm) prevents z-fighting on surfaces to which AR object is "pressed".

Soft Occlusion: Blurred Edges

At LiDAR mesh edges occlusion transition sometimes "pixelated" — depth resolution finite. Soft occlusion via bilateral blur on depth mask smooths transition. In Metal: custom kernel with depth-aware blur (blur only within depth threshold).

Noticeable only on careful look, but for premium AR product worth implementing.

Shadows from Virtual Objects on Real Surfaces

Bonus to occlusion: AR sofa casts shadow on real floor. In RealityKit with sceneUnderstanding.receivesLighting — shadows automatic. In custom render: shadow map from AR object, projected on real-world geometry from LiDAR mesh.

Without LiDAR: shadow only on ARKit-detected planes (floor, table). User sees "flat" shadow not following real surface relief.

Timeline

LiDAR occlusion via RealityKit API — 2–3 days integration. Person segmentation occlusion — 3–5 days. ARCore Depth occlusion on Android — 1–2 weeks (custom shader). Soft occlusion with custom Metal/GLSL shading — plus 1–2 weeks. Cost calculated individually.