New SEPatch3D framework accelerates Vision Transformer-based 3D object detectors by dynamically adjusting patch sizes while preserving semantic information.

Existing token compression methods (pruning, merging, patch enlargement) inadvertently discard background information and disrupt contextual consistency in 3D detection tasks

SEPatch3D introduces Spatiotemporal-aware Patch Size Selection (SPSS) that dynamically assigns smaller patches to scenes with nearby objects for detail preservation and larger patches to background-dominated areas for computational efficiency

The approach addresses key limitations of previous acceleration strategies by maintaining fine-grained semantics while reducing inference latency in ViT-based sparse multi-view 3D object detectors