.png){kind=icon}
.png){kind=icon}
Return
DECAD: Diffusion-Based Enhanced Condition Encoder for Skeleton Anomaly Detection
.png){kind=icon}
PRE
AI
DOI：10.1109/lsp.2026.3723342.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffusion-based skeleton anomaly detection has achieved promising accuracy, yet prevailing methods rely on increasingly heavyweight condition encoders—some even larger than the diffusion backbones they condition—hindering deployment on resource-constrained platforms. We argue that effective conditioning does not require such capacity inflation. We propose DECAD, a parameter-efficient enhanced condition encoder built upon two lightweight modules. The Structure-Aware Graph Fusion (SAGF) module fuses multi-head self-attention with graph convolution to capture both anatomy-fixed and motion-adaptive joint relationships. The Adaptive Channel–Spatial Attention (ACSA) module then progressively recalibrates conditional features via channel attention, channel shuffle, and spatial attention. With merely 10.34 K encoder parameters, DECAD achieves competitive detection accuracy while reducing total model parameters by over two orders of magnitude and accelerating end-to-end inference by 1.56× over prevailing diffusion-based approaches. Extensive experiments on four public benchmarks validate that lightweight conditional encoding suffices for skeleton-based anomaly detection.
Keywords:
Skeleton-based anomaly detection
diffusion model
condition encoder
Journal
I
IF:
3.9
Papers:
610
Citations:
0