GTR: Gated Token Recurrence for Efficient Dense Prediction

2026-09-22

AI Summary

为解决自注意力机制在高分辨率图像上的效率问题，本文提出Gated Token Recurrence (GTR)方法，采用无softmax的循环结构，提高了密集预测任务的效率。

Abstract

Self-attention-based vision backbones perform well on dense prediction, but the quadratic computational cost of global softmax attention limits their efficiency as image resolution increases. We introduce Gated Token Recurrence (GTR), a softmax-free recurrent vision backbone that combines gated linear attention, alternating spatial scan directions, and spatially enhanced SwiGLU blocks. GTR is distilled from a detection-specialized DINOv3 teacher using only final-layer patch-token alignment through a linear projection and squared $\ell_2$ loss, without masked-token prediction or intermediate-layer supervision. With Objects365 detector pre-training, GTR-L achieves 58.9 box AP on COCO \texttt{val2017} with 1.908\,ms median batch-one latency under compiled FP16 execution on an RTX~4090. The same backbone also transfers to instance segmentation, pose estimation, oriented detection, semantic segmentation, and monocular depth estimation. In an isolated kernel benchmark, our specialized chunkwise CUDA operator is $4.0\times$ faster than FLA v0.5.0 at 1.6K tokens on RTX~4090. TensorRT deployment on DRIVE AGX Thor achieves 2.282--8.769\,ms median batch-one latency across the evaluated models. These results show that recurrent token mixing can provide an efficient alternative to global softmax attention for high-resolution dense prediction and edge deployment.Project page: https://intellindust-ai-lab.github.io/projects/GTR/

Problem

• self-attention
• dense prediction
• computational cost
• image resolution
• efficiency

Innovation

• Gated Token Recurrence
• gated linear attention
• spatially enhanced SwiGLU blocks
• efficient dense prediction