I-LW-DETR enables integer-only inference for detection transformers

Deploying vision transformer detectors on edge devices like NPUs and microcontrollers has been a challenge because key operations—deformable attention, feature fusion, Softmax, GELU, and LayerNorm—do not run natively in integer arithmetic. Existing quantized solutions either retain floating-point operators or focus on heavyweight backbones, leaving lightweight detection transformers without an end-to-end integer pipeline. Researchers from a collaboration (authors Thanh Cong Le, Michal Szczepanski, Martyna Poreba) propose I-LW-DETR, the first fully integer-only lightweight DETR. Their approach introduces three components: a scale-preserving split convolution that assigns independent activation scales to each branch of the multi-scale projector; SD-ShiftGELU, a sign-dependent approximation of GELU that preserves element-wise behavior; and a constrained Shiftmax that maintains stable Softmax normalization.

Experimental results across multiple model scales demonstrate that I-LW-DETR achieves a 3.6x reduction in model size and reduces computational cost by more than one order of magnitude (10x+), with only moderate accuracy degradation. This breakthrough makes it feasible to run vision transformer detectors on hardware that lacks floating-point units, opening the door to real-time object detection on low-power edge devices. The paper is available on arXiv (2607.24981).