library_name: pytorch
license: other
tags:
- android
pipeline_tag: depth-estimation
Midas-V2: Optimized for Mobile Deployment
Deep Convolutional Neural Network model for depth estimation
Midas is designed for estimating depth at each point in an image.
This model is an implementation of Midas-V2 found here.
This repository provides scripts to run Midas-V2 on Qualcomm® devices. More details on model performance across various devices, can be found here.
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): 16.6 MB
| Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model |
|---|---|---|---|---|---|---|---|---|
| Midas-V2 | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 12.847 ms | 0 - 42 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 11.829 ms | 1 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 4.969 ms | 0 - 50 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 6.483 ms | 0 - 37 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 3.206 ms | 0 - 307 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 2.973 ms | 0 - 12 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 4.48 ms | 0 - 42 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 4.162 ms | 1 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 12.847 ms | 0 - 42 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 11.829 ms | 1 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 3.229 ms | 0 - 288 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 2.978 ms | 0 - 26 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 5.889 ms | 0 - 28 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 5.331 ms | 1 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 3.228 ms | 0 - 311 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 2.978 ms | 0 - 13 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 4.48 ms | 0 - 42 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 4.162 ms | 1 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 3.236 ms | 0 - 310 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN_DLC | 2.987 ms | 1 - 12 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 2.957 ms | 0 - 73 MB | NPU | Midas-V2.onnx |
| Midas-V2 | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 2.351 ms | 0 - 67 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 2.075 ms | 1 - 43 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 2.034 ms | 0 - 49 MB | NPU | Midas-V2.onnx |
| Midas-V2 | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 1.787 ms | 0 - 44 MB | NPU | Midas-V2.tflite |
| Midas-V2 | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN_DLC | 1.887 ms | 0 - 33 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 1.924 ms | 1 - 32 MB | NPU | Midas-V2.onnx |
| Midas-V2 | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 3.543 ms | 180 - 180 MB | NPU | Midas-V2.dlc |
| Midas-V2 | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 3.052 ms | 35 - 35 MB | NPU | Midas-V2.onnx |
| Midas-V2 | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 2.453 ms | 0 - 28 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN_DLC | 2.844 ms | 0 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 1.443 ms | 0 - 51 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN_DLC | 1.821 ms | 0 - 47 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 1.063 ms | 0 - 148 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN_DLC | 1.285 ms | 0 - 133 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 1.404 ms | 0 - 29 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN_DLC | 1.583 ms | 0 - 30 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | TFLITE | 3.748 ms | 0 - 46 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | RB3 Gen 2 (Proxy) | Qualcomm® QCS6490 (Proxy) | QNN_DLC | 5.733 ms | 0 - 46 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | RB5 (Proxy) | Qualcomm® QCS8250 (Proxy) | TFLITE | 15.948 ms | 0 - 2 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 2.453 ms | 0 - 28 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA7255P ADP | Qualcomm® SA7255P | QNN_DLC | 2.844 ms | 0 - 29 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 1.075 ms | 0 - 148 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN_DLC | 1.292 ms | 0 - 133 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 1.903 ms | 0 - 31 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA8295P ADP | Qualcomm® SA8295P | QNN_DLC | 2.191 ms | 0 - 33 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 1.076 ms | 0 - 150 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN_DLC | 1.274 ms | 0 - 133 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 1.404 ms | 0 - 29 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | SA8775P ADP | Qualcomm® SA8775P | QNN_DLC | 1.583 ms | 0 - 30 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 1.07 ms | 0 - 149 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN_DLC | 1.296 ms | 0 - 133 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 129.801 ms | 13 - 126 MB | NPU | Midas-V2.onnx |
| Midas-V2 | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 0.761 ms | 0 - 58 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN_DLC | 0.906 ms | 0 - 58 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 100.155 ms | 23 - 483 MB | NPU | Midas-V2.onnx |
| Midas-V2 | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 0.718 ms | 0 - 29 MB | NPU | Midas-V2.tflite |
| Midas-V2 | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN_DLC | 0.785 ms | 0 - 31 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 89.688 ms | 13 - 454 MB | NPU | Midas-V2.onnx |
| Midas-V2 | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN_DLC | 1.504 ms | 119 - 119 MB | NPU | Midas-V2.dlc |
| Midas-V2 | w8a8 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 144.558 ms | 69 - 69 MB | NPU | Midas-V2.onnx |
Installation
Install the package via pip:
pip install "qai-hub-models[midas]"
Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
Sign-in to Qualcomm® AI Hub with your
Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.
With this API token, you can configure your client to run models on the cloud hosted devices.
qai-hub configure --api_token API_TOKEN
Navigate to docs for more information.
Demo off target
The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.
python -m qai_hub_models.models.midas.demo
The above demo runs a reference implementation of pre-processing, model inference, and post processing.
NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).
%run -m qai_hub_models.models.midas.demo
Run model on a cloud-hosted device
In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:
- Performance check on-device on a cloud-hosted device
- Downloads compiled assets that can be deployed on-device for Android.
- Accuracy check between PyTorch and on-device outputs.
python -m qai_hub_models.models.midas.export
Profiling Results
------------------------------------------------------------
Midas-V2
Device : cs_8275 (ANDROID 14)
Runtime : TFLITE
Estimated inference time (ms) : 12.8
Estimated peak memory usage (MB): [0, 42]
Total # Ops : 138
Compute Unit(s) : npu (138 ops) gpu (0 ops) cpu (0 ops)
How does this work?
This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:
Step 1: Compile model for on-device deployment
To compile a PyTorch model for on-device deployment, we first trace the model
in memory using the jit.trace and then call the submit_compile_job API.
import torch
import qai_hub as hub
from qai_hub_models.models.midas import Model
# Load the model
torch_model = Model.from_pretrained()
# Device
device = hub.Device("Samsung Galaxy S24")
# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
# Compile model on a specific device
compile_job = hub.submit_compile_job(
model=pt_model,
device=device,
input_specs=torch_model.get_input_spec(),
)
# Get target model to run on-device
target_model = compile_job.get_target_model()
Step 2: Performance profiling on cloud-hosted device
After compiling models from step 1. Models can be profiled model on-device using the
target_model. Note that this scripts runs the model on a device automatically
provisioned in the cloud. Once the job is submitted, you can navigate to a
provided job URL to view a variety of on-device performance metrics.
profile_job = hub.submit_profile_job(
model=target_model,
device=device,
)
Step 3: Verify on-device accuracy
To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.
input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
model=target_model,
device=device,
inputs=input_data,
)
on_device_output = inference_job.download_output_data()
With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.
Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.
Run demo on a cloud-hosted device
You can also run the demo on-device.
python -m qai_hub_models.models.midas.demo --eval-mode on-device
NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).
%run -m qai_hub_models.models.midas.demo -- --eval-mode on-device
Deploying compiled model to Android
The models can be deployed using multiple runtimes:
TensorFlow Lite (
.tfliteexport): This tutorial provides a guide to deploy the .tflite model in an Android application.QNN (
.soexport ): This sample app provides instructions on how to use the.soshared library in an Android application.
View on Qualcomm® AI Hub
Get more details on Midas-V2's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of Midas-V2 can be found here.
- The license for the compiled assets for on-device deployment 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.