library_name: onnx
license: other
tags:
- foundation
- amd
- rocm
- depth-estimation
pipeline_tag: depth-estimation
Bi3D: Optimized for AMD ROCm
Bi3D performs stereo depth estimation by reformulating disparity search as a sequence of binary classifications over a cost volume, with optional 3D regularization for sub-pixel continuous depth. This repository packages evaluation/inference for stereo depth / disparity estimation using ONNX Runtime, MIGraphX EP, and VitisAI EP, exported and validated for AMD ROCm so it runs efficiently on AMD GPUs, CPUs, and NPUs.
This is based on the implementation of Bi3D found here. This repository contains configurations and scripts optimized for AMD® ROCm™ platforms. You can use the Bi3D AMD scripts to reproduce results or export with custom configurations.
Task Overview
Task: Stereo depth / disparity estimation
Dataset: SceneFlow FlyingThings3D TEST (cleanpass WebP; bundled 10-pair subset committed in-repo)
Output metrics: EPE (End-Point Error, mean absolute disparity error in pixels)
NPU note: Uses VitisAI EP auto-partitioning with per-dtype
config/vitisai_config_*.jsonfiles.
AMD ROCm Optimization
This model export has been adapted and validated for AMD Instinct™ / Radeon™ GPUs running ROCm, as well as AMD CPUs and NPUs. Key points:
- Validated backends: ONNX Runtime (CPU), MIGraphX EP (GPU), and VitisAI EP (NPU).
| Runtime | Precision | Backend | Hardware | Notes |
|---|---|---|---|---|
| CPU | FP32 / FP16 / BF16 | ONNX Runtime | AMD RYZEN AI MAX+ 395 w/ Radeon 8060S | — |
| GPU | FP32 / FP16 / BF16 | MIGraphX EP | AMD RYZEN AI MAX+ 395 w/ Radeon 8060S | — |
| NPU | FP32 / FP16 / BF16 | VitisAI EP | AMD RYZEN AI MAX+ 395 w/ Radeon 8060S | Auto-partitioning via per-dtype config/vitisai_config_*.json |
Getting Started
For setup instructions, evaluation scripts, and custom configuration options, see the Bi3D on GitHub.
Model Details
Model Type: Stereo depth estimation (binary classification over a disparity cost volume, with 3D regularization for continuous depth)
Base Model: SceneFlow-trained continuous depth 3D + confidence regularization checkpoint (NVIDIA Bi3D)
Model Stats:
- Model variant: continuous depth 3D + confidence regularization (default checkpoint)
- Input resolution: 288×480 (cropped from the original 540×960 to fit device memory constraints)
- Precision tested: FP32, FP16, BF16
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