SCARF: A Scene-Adaptive Depth-Guided G-3DGS Encoder Accelerator with Semantic Reuse and Fused Dataflow

Zirui Ma · Zhihua Fan · Haibin Wu · Wenxing Li · Fulin Zhang · Dongrui Fan · Xiaochun Ye · Wenming Li

Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences

MICRO 2026

ACM Artifacts Available badgeAvailable
ACM Artifacts Evaluated Functional badgeFunctional
ACM Results Reproduced badgeReproduced
Paper Code Artifact Talk
Bedroom scene used for Gaussian similarity analysis in the paper
Scene with one example tile
Encoder features · schematic
Query Q · 16-bit LSH Hamming distance ≤ threshold Cached C · 16-bit LSH Cached depth anchor dc
Same image pixels · corner probes
Semantic match · cached depth
Depth candidates · same global set
BaselineD
Cache hitD/R
A miss retains all D candidates.
Cross-view sampling → cost volume → U-Net → depth
4 × 4 TILE · CORNER PROBESOriginal S2 depth + Gaussian adaptor
L0Low feature varianceRepresentative
L1Else, low depth spreadDepth-weighted
FullElse, irregular tilePer-pixel adaptor
Only non-probe pixels follow the selected path.

Abstract

Depth-guided G-3DGS encoders reconstruct scenes directly from multi-view images without per-scene optimization. SCARF addresses two sources of redundant encoder work: FSDR reuses depth information across semantically similar features, while SAES selects a Gaussian generation path from local scene regularity before unnecessary primitives are created. A fused accelerator implements both mechanisms and delivers a 2.94-fold geometric-mean end-to-end speedup on the evaluated workloads, with maximum PSNR degradation below 0.072 percent.

Dataflow and Architecture

Complete SCARF dataflow: feature extraction, FSDR depth search, probe-driven SAES materialization, and Gaussian conversion.

FSDR: Feature Similarity Depth Reuse

FSDR matches a query signature against the semantic cache. A hit selects D/R original candidates around the cached depth; a miss retains all D. Both paths use the original depth predictor and update the cache.

SAES: Scene-Adaptive Early Sparsification

SAES computes four corner probes with the original depth predictor and Gaussian adaptor. Non-probe pixels use L0 soft assignment and moment matching for low feature variance, otherwise L1 depth-weighted aggregation for low depth spread, otherwise full per-pixel computation.
SCARF architecture with FSDR cache, matrix-vector unit, shared compute array, local buffers, pipeline controller, and Gaussian generation unit.

Qualitative Visualization

Quality change
FSDR worst-view quality change
Stage latency
FSDR worst-view stage latency
Rendered view
FSDR worst-view rendered view
RGB error
FSDR worst-view RGB errorOriginal RGB error color scale

Quantitative Comparison

Faster end to end.

End-to-end speedup over Orin NX for nine model-dataset pairs and normalized stage latency breakdown.

Presentation

Opening slide of the SCARF conference presentation videoXIAOHONGSHU ↗

Artifact

Paper Source code Reproduction guide Archived artifact

BibTeX

@inproceedings{ma2026scarf,
  author    = {Zirui Ma and Zhihua Fan and Haibin Wu and Wenxing Li and
               Fulin Zhang and Dongrui Fan and Xiaochun Ye and Wenming Li},
  title     = {{SCARF}: A Scene-Adaptive Depth-Guided {G-3DGS}
               Encoder Accelerator with Semantic Reuse and Fused Dataflow},
  booktitle = {2026 59th IEEE/ACM International Symposium
               on Microarchitecture (MICRO)},
  year      = {2026},
  pages     = {2512--2528},
  publisher = {IEEE},
  doi       = {10.1109/MICRO67795.2026.00172},
  url       = {https://doi.org/10.1109/MICRO67795.2026.00172}
}