raft-stereo

RAFT-Stereo: Optimized for AMD ROCm

RAFT-Stereo (Princeton VL) estimates dense per-pixel disparity from a rectified stereo image pair using multilevel recurrent field transforms. This repository packages evaluation/inference for stereo disparity estimation using ONNX Runtime, exported and validated for AMD ROCm so it runs efficiently on AMD GPUs and CPUs, as well as AMD Ryzen AI NPUs.

This is based on the implementation of RAFT-Stereo found here. This repository contains configurations and scripts optimized for AMD® ROCm™ platforms. You can use the RAFT-Stereo AMD scripts to reproduce results or export with custom configurations.

---

Task Overview

Task: Stereo disparity estimation

Dataset: Middlebury MiddEval3 (F/H/Q splits), ETH3D two-view training set, KITTI 2012/2015 stereo, and FlyingThings3D (SceneFlow)

Output metrics: EPE (End-Point Error, px), D1-error (%)

ONNX Runtime note: CPU runs FP32 only. GPU supports FP32/FP16/INT8 individually or all at once via the MIGraphX execution provider. NPU has no precision knob (Vitis AI compiles the graph as-is); instead it offers a smoke graph (few unrolled GRU iterations, compiles in minutes) vs. the full graph (32 iterations, Vitis AI compilation can take hours).

---

AMD ROCm Optimization

This model export has been adapted and validated for AMD Instinct™ / Radeon™ GPUs running ROCm, as well as AMD CPUs and AMD Ryzen AI NPUs. Key points:

• Exported/tested with ROCm 7.x for the MIGraphX execution provider.
• Validated backends: ONNX Runtime CPU EP, ONNX Runtime MIGraphX EP (ROCm GPU), and Vitis AI EP (Ryzen AI NPU).
• No code changes required versus the upstream RAFT-Stereo implementation — only environment/runtime configuration differs.
• CPU fallback path supported for environments without a ROCm-capable GPU.

---

Getting Started

For setup instructions, evaluation scripts, and custom configuration options, see the RAFT-Stereo on GitHub.

---

Model Details

Model Type: Stereo disparity estimation (multilevel recurrent field transform network)

Model Stats:

• No MODEL_SIZE variants — single architecture
• Static ONNX export by default: [1, 3, 384, 512] input resolution (configurable via --height/--width, or --dynamic for variable axes)
• Precision tested: FP32 (CPU); FP32, FP16, INT8 (GPU); Vitis AI-compiled (NPU)

---

Accuracy Pipeline

Accuracy evaluation is not yet implemented for this model.

---

Dig Deeper

Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?

📂 View the full project on GitHub

The GitHub repository includes:

• Setup and prerequisites for ROCm environments
• Scripts for the supported runners
• Additional model variants and datasets
• Benchmarking and reproduction instructions