Scheduling Level of Detail with guaranteed Quality and Cost W. Pasman, F. W. Jansen Delft University of Technology

UbiCom mobile AR

Latency Layering Optical AR -> Low latency (10ms target for our system) on low-power headset -> Latency layering

Low-Latency Rendering Render just ahead of rasterbeam Voodoo2 3D game card measured latency 8.5ms

Dynamic Simplification Dynamic LoD generation in backbone Maximize perf/cost ratio in headset.

Mathematical model Estimate link and CPU load, memory usage, lifetime of objects, etc Est screenspace error derived from geometric distortions D=0.001 R=1m

Overview Mathematical model only for single objects -> complex scenes require resource scheduling Accuracy Curves VRML Integration Measurements

Scheduling of resources Funkhouser,Séquin (1992), Mason,Blake (1997) Goal: maximize benefit within a cost budget.

Mason showed this is NP complete.. - Iterative approx giving in worst case half the maximum possible quality - Quality only known after iteration - Only feedback loop with application possible

Accuracy Curves

Each node in scene graph is assigned a curve: Accuracy curve required resources as function of accuracy target monotonically increasing. R->#polygons

Measurement of geometric distortion d as function of number of polygons n d~ C/n. Accuracy a =1/d Resource usage r = K a + R 0 -> piecewise linear function

Propagating accuracy curves Leaf nodes: accuracy curve from (1) mathematical model or (2) measurements Other nodes: propagate curve upwards through scenegraph

Resource Usage LoD node behaviour

Different end points curve 1 curve 2 Accuracy Resource usage incorrect extension incorrect minimum 2050

Complexity Target Accuracy -> Resources required Idem R -> A (so not NP-complete)

Optimizing curve updates Upto now: screenspace error=visual accuracy Refresh required if user moves Optimizations: (1) Determine range where a curve is ‘accurate enough’ as long as viewer is within the range. (2) Visual accuracy is derived from geometric distortion - which is viewpoint independent Relative acc = geometric accuracy object radius

Using relative accuracy curves Slight changes in algorithms: grouping -> 'object' diameter changes. Conversion to visual accuracy needed. Group's bbox is much closer to viewer than the individual bboxes -> switch to absolute accuracy at reasonable distance.

VRML Integration

ConicRange Accuracy curves & simplified objects: valid only in part of space. Efficient checking. d

Imposter nodes Replace children with image. Automatic refresh & calc of accuracy curve SimpleImposter { MFNodechildren[ ] SFVec2fsize2 2 SFBoolautosizetrue SFVec3fcenter0 0 0 SFBoolautocentertrue }

LOD LOD picks valid level requiring least resources

Measurements

Prototype limitations -> simple testscene Target #polygons, accuracy Measure also SNR

Constant polygon budget Constant accuracy target

Conclusions Efficient scheduling of resources Fits all simplification methods Rendering to target quality or budget driven Comes close to budget, never exceeds it. Hierarchical approach fits with VRML Many optimization possibilities

Future work Limited prototype: part of UbiCom AR system Many representations can change at once. This causes peak datarates. Extension to multiple resources Conic ranges quickly degenerate Animated objects