Pip-Stereo：面向迭代优化立体匹配的渐进式迭代剪枝器
Pip-Stereo: Progressive Iterations Pruner for Iterative Optimization based Stereo Matching
把迭代式立体匹配的递归细化过程按空间稀疏性与时间冗余性做渐进式迭代剪枝，并用硬件感知的 FlashGRU 替代 ConvGRU，为在边缘设备上部署迭代式匹配网络提供了可复用的加速范式。
原文依据 we develop FlashGRU, a hardware-aware RNN operator leveraging structured sparsity and I/O-conscious design, achieving a 7.28xspeedup
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论迭代式立体匹配虽能达到较高精度，但其对循环神经网络（RNN）的依赖阻碍了边缘端部署，这一挑战在现有研究中尚未得到充分探索。我们分析了迭代细化过程，发现视差更新在空间上是稀疏的、在时间上是冗余的。首先，我们引入一种渐进式迭代剪枝策略，抑制冗余的更新步骤，有效地将递归计算压缩为近乎单次前向推理。其次，我们提出一种协同单目先验迁移框架，无需专用的单目编码器即可隐式嵌入深度先验，从而消除其带来的计算负担。第三，我们设计了 FlashGRU，一种利用结构化稀疏性与 I/O 感知设计的硬件感知 RNN 算子，在 2K 分辨率下相比原生 ConvGRU 实现 7.28 倍加速、76.6% 的显存峰值降低以及 80.9% 的全局内存请求减少。我们的 Pip-Stereo 使边缘硬件上的实时高保真立体匹配成为可能：在 NVIDIA Jetson Orin NX（FP16）上处理 320x640 帧仅需 75ms，在 RTX 4090 上仅需 19ms，精度可媲美大型迭代式模型，且我们的泛化能力与精度远超现有实时方法。
英文原文摘要 1270 字符
While iterative stereo matching achieves high accuracy, its dependence on Recurrent Neural Networks (RNN) hinders edge deployment, a challenge underexplored in existing researches. We analyze iterative refinement and reveal that disparity updates are spatially sparse and temporally redundant. First, we introduce a progressive iteration pruning strategy that suppresses redundant update steps, effectively collapsing the recursive computation into a near-single-pass inference. Second, we propose a collaborative monocular prior transfer framework that implicitly embeds depth priors without requiring a dedicated monocular encoder, thereby eliminating its associated computational burden. Third, we develop FlashGRU, a hardware-aware RNN operator leveraging structured sparsity and I/O-conscious design, achieving a 7.28xspeedup, 76.6% memory peak reduction and 80.9% global memory requests reduction over natvie ConvGRUs under 2K resolution. Our PipStereo enables real-time, high-fidelity stereo matching on edge hardware: it processes 320x640 frames in just 75ms on an NVIDIA Jetson Orin NX (FP16) and 19ms on RTX 4090, matching the accuracy of large iterative based models, and our generalization ability and accuracy far exceeds that of existing real-time methods.