Midas-V2: Optimized for Qualcomm Devices
Midas is designed for estimating depth at each point in an image.
This is based on the implementation of Midas-V2 found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.
Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up 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 |
| ONNX | w8a8 | Universal | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| QNN_DLC | float | Universal | QAIRT 2.45 | Download |
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | Download |
| TFLITE | float | Universal | QAIRT 2.45 | Download |
| TFLITE | w8a8 | Universal | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit Midas-V2 on Qualcomm® AI Hub.
Option 2: Export with Custom Configurations
Use the Qualcomm® AI Hub Models 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 Midas-V2 on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.depth_estimation
Model Stats:
- Model checkpoint: MiDaS_small
- Input resolution: 256x256
- Number of parameters: 16.6M
- Model size (float): 63.2 MB
- Model size (w8a8): 16.9 MB
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| Midas-V2 | ONNX | float | Snapdragon® X2 Elite | 1.314 ms | 2 - 2 MB | NPU |
| Midas-V2 | ONNX | float | Snapdragon® X Elite | 2.774 ms | 36 - 36 MB | NPU |
| Midas-V2 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 1.932 ms | 0 - 198 MB | NPU |
| Midas-V2 | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 6.48 ms | 1 - 199 MB | NPU |
| Midas-V2 | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 2.725 ms | 0 - 72 MB | NPU |
| Midas-V2 | ONNX | float | Qualcomm® QCS8450 | 6.48 ms | 1 - 199 MB | NPU |
| Midas-V2 | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 3.652 ms | 1 - 4 MB | NPU |
| Midas-V2 | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 2.774 ms | 36 - 36 MB | NPU |
| Midas-V2 | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 1.475 ms | 0 - 163 MB | NPU |
| Midas-V2 | ONNX | float | Snapdragon® 8 Elite Mobile | 1.475 ms | 0 - 163 MB | NPU |
| Midas-V2 | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.26 ms | 0 - 164 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® X2 Elite | 0.616 ms | 1 - 1 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® X Elite | 1.332 ms | 19 - 19 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.925 ms | 0 - 96 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.838 ms | 0 - 96 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 4.719 ms | 0 - 3 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.327 ms | 0 - 22 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® QCS8450 | 1.838 ms | 0 - 96 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.824 ms | 0 - 3 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 1.332 ms | 19 - 19 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.761 ms | 0 - 59 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® 8 Elite Mobile | 0.761 ms | 0 - 59 MB | NPU |
| Midas-V2 | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.656 ms | 0 - 61 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® X2 Elite | 1.583 ms | 1 - 1 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® X Elite | 3.273 ms | 1 - 1 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 2.116 ms | 0 - 193 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 6.838 ms | 0 - 192 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8275 | 12.104 ms | 1 - 158 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.033 ms | 0 - 5 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® SA8775P | 4.242 ms | 1 - 160 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® SA8650P | 4.242 ms | 1 - 160 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® SA8255P | 4.242 ms | 1 - 160 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® QCS8450 | 6.838 ms | 0 - 192 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 4.544 ms | 3 - 5 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 3.273 ms | 1 - 1 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 1.602 ms | 1 - 163 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® SA7255P | 12.104 ms | 1 - 158 MB | NPU |
| Midas-V2 | QNN_DLC | float | Qualcomm® SA8295P | 5.445 ms | 1 - 155 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 1.602 ms | 1 - 163 MB | NPU |
| Midas-V2 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.293 ms | 1 - 163 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.728 ms | 0 - 0 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® X Elite | 1.445 ms | 0 - 0 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.925 ms | 0 - 86 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.906 ms | 0 - 85 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 5.633 ms | 0 - 2 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8275 | 2.895 ms | 0 - 49 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.322 ms | 0 - 2 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® SA8775P | 1.594 ms | 0 - 51 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® SA8650P | 1.594 ms | 0 - 51 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® SA8255P | 1.594 ms | 0 - 51 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 1.906 ms | 0 - 85 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.447 ms | 0 - 2 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 1.445 ms | 0 - 0 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 9.149 ms | 0 - 175 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.606 ms | 0 - 172 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.697 ms | 0 - 47 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® SA7255P | 2.895 ms | 0 - 49 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Qualcomm® SA8295P | 2.253 ms | 0 - 48 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 0.697 ms | 0 - 47 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.578 ms | 0 - 52 MB | NPU |
| Midas-V2 | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.606 ms | 0 - 172 MB | NPU |
| Midas-V2 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 2.119 ms | 0 - 226 MB | NPU |
| Midas-V2 | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 6.809 ms | 0 - 215 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® Dragonwing™ QCS8275 | 12.149 ms | 0 - 172 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 3.027 ms | 0 - 2 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® SA8775P | 4.263 ms | 0 - 173 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® SA8650P | 4.263 ms | 0 - 173 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® SA8255P | 4.263 ms | 0 - 173 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® QCS8450 | 6.809 ms | 0 - 215 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 4.017 ms | 0 - 39 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 1.602 ms | 0 - 174 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® SA7255P | 12.149 ms | 0 - 172 MB | NPU |
| Midas-V2 | TFLITE | float | Qualcomm® SA8295P | 5.486 ms | 0 - 162 MB | NPU |
| Midas-V2 | TFLITE | float | Snapdragon® 8 Elite Mobile | 1.602 ms | 0 - 174 MB | NPU |
| Midas-V2 | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 1.299 ms | 0 - 175 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.754 ms | 0 - 84 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.598 ms | 0 - 84 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 5.304 ms | 0 - 29 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8275 | 2.508 ms | 0 - 46 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.082 ms | 0 - 85 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® SA8775P | 1.374 ms | 0 - 50 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® SA8650P | 1.374 ms | 0 - 50 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® SA8255P | 1.374 ms | 0 - 50 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® QCS8450 | 1.598 ms | 0 - 84 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.223 ms | 0 - 20 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 8.407 ms | 0 - 179 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.377 ms | 0 - 169 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.6 ms | 0 - 52 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® SA7255P | 2.508 ms | 0 - 46 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Qualcomm® SA8295P | 2.026 ms | 0 - 45 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 0.6 ms | 0 - 52 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.499 ms | 0 - 49 MB | NPU |
| Midas-V2 | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.377 ms | 0 - 169 MB | NPU |
License
- The license for the original implementation of Midas-V2 can be found here.
References
- Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer
- Source Model Implementation
Community
- Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
- For questions or feedback please reach out to us.
