README.md

---
library_name: pytorch
license: other
tags:
- android
pipeline_tag: depth-estimation

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/crestereo/web-assets/model_demo.png)

# CREStereo: Optimized for Qualcomm Devices

CREStereo (Cascaded Recurrent Network with Adaptive Correlation) is a CVPR 2022 Oral paper that achieves state-of-the-art stereo matching accuracy.

This is based on the implementation of CREStereo found [here](https://github.com/ibaiGorordo/CREStereo-Pytorch).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.60.0/src/qai_hub_models/models/crestereo) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.45, ONNX Runtime 1.27.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/crestereo/releases/v0.60.0/crestereo-onnx-float.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/crestereo/releases/v0.60.0/crestereo-qnn_dlc-float.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/crestereo/releases/v0.60.0/crestereo-tflite-float.zip)

For more device-specific assets and performance metrics, visit **[CREStereo on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/crestereo)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.60.0/src/qai_hub_models/models/crestereo) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [CREStereo on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.60.0/src/qai_hub_models/models/crestereo) for usage instructions.

## Model Details

**Model Type:** Model_use_case.depth_estimation

**Model Stats:**
- Input: Rectified stereo pair — left and right RGB images
- Input resolution: 240x320
- Model checkpoint: CREStereo ETH3D pretrained (crestereo_eth3d.pt)
- Model size (float): 20.7 MB
- Number of parameters: 5.43M
- Output: Disparity map

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| CREStereo | ONNX | float | Snapdragon® X2 Elite | 87.564 ms | 2 - 2 MB | NPU
| CREStereo | ONNX | float | Snapdragon® X Elite | 202.736 ms | 21 - 21 MB | NPU
| CREStereo | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 144.369 ms | 3 - 1334 MB | NPU
| CREStereo | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 328.168 ms | 3 - 1366 MB | NPU
| CREStereo | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 271.157 ms | 2 - 6 MB | NPU
| CREStereo | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 200.076 ms | 0 - 25 MB | NPU
| CREStereo | ONNX | float | Qualcomm® QCS8450 | 328.168 ms | 3 - 1366 MB | NPU
| CREStereo | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 269.024 ms | 2 - 5 MB | NPU
| CREStereo | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 202.736 ms | 21 - 21 MB | NPU
| CREStereo | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 111.31 ms | 1 - 1112 MB | NPU
| CREStereo | ONNX | float | Snapdragon® 8 Elite Mobile | 111.31 ms | 1 - 1112 MB | NPU
| CREStereo | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 79.199 ms | 1 - 837 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® X2 Elite | 88.18 ms | 2 - 2 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® X Elite | 203.778 ms | 2 - 2 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 144.84 ms | 2 - 1242 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 328.828 ms | 2 - 1292 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 271.144 ms | 2 - 7 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 797.535 ms | 3 - 1015 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 199.896 ms | 3 - 6 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® SA8775P | 269.996 ms | 3 - 1059 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® SA8650P | 269.996 ms | 3 - 1059 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® SA8255P | 269.996 ms | 3 - 1059 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® QCS8450 | 328.828 ms | 2 - 1292 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 269.305 ms | 4 - 8 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 203.778 ms | 2 - 2 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 111.55 ms | 3 - 1046 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® SA7255P | 797.535 ms | 3 - 1015 MB | NPU
| CREStereo | QNN_DLC | float | Qualcomm® SA8295P | 312.046 ms | 3 - 1066 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 111.55 ms | 3 - 1046 MB | NPU
| CREStereo | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 79.208 ms | 1 - 748 MB | NPU
| CREStereo | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 151.227 ms | 1 - 1672 MB | NPU
| CREStereo | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 379.427 ms | 1 - 1643 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 278.444 ms | 0 - 44 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 809.84 ms | 1 - 1402 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 216.12 ms | 0 - 7 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® SA8775P | 282.462 ms | 1 - 1411 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® SA8650P | 282.462 ms | 1 - 1411 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® SA8255P | 282.462 ms | 1 - 1411 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® QCS8450 | 379.427 ms | 1 - 1643 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 281.541 ms | 0 - 43 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 122.089 ms | 1 - 1377 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® SA7255P | 809.84 ms | 1 - 1402 MB | NPU
| CREStereo | TFLITE | float | Qualcomm® SA8295P | 350.245 ms | 1 - 1383 MB | NPU
| CREStereo | TFLITE | float | Snapdragon® 8 Elite Mobile | 122.089 ms | 1 - 1377 MB | NPU
| CREStereo | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 88.324 ms | 0 - 1114 MB | NPU

## License
* The license for the original implementation of CREStereo can be found
  [here](https://github.com/megvii-research/CREStereo/blob/master/LICENSE).

## References
* [Practical Stereo Matching via Cascaded Recurrent Network with Adaptive Correlation](https://arxiv.org/abs/2203.11483)
* [Source Model Implementation](https://github.com/ibaiGorordo/CREStereo-Pytorch)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).