--- library_name: pytorch license: apache-2.0 pipeline_tag: object-detection tags: - real_time - android --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/mediapipe_face/web-assets/model_demo.png) # MediaPipe-Face-Detection: Optimized for Mobile Deployment ## Detect faces and locate facial features in real-time video and image streams Designed for sub-millisecond processing, this model predicts bounding boxes and pose skeletons (left eye, right eye, nose tip, mouth, left eye tragion, and right eye tragion) of faces in an image. This model is an implementation of MediaPipe-Face-Detection found [here](https://github.com/zmurez/MediaPipePyTorch/). This repository provides scripts to run MediaPipe-Face-Detection on Qualcomm® devices. More details on model performance across various devices, can be found [here](https://aihub.qualcomm.com/models/mediapipe_face). ### Model Details - **Model Type:** Object detection - **Model Stats:** - Input resolution: 256x256 - Number of parameters (MediaPipeFaceDetector): 135K - Model size (MediaPipeFaceDetector): 565 KB - Number of parameters (MediaPipeFaceLandmarkDetector): 603K - Model size (MediaPipeFaceLandmarkDetector): 2.34 MB - Number of output classes: 6 | Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model |---|---|---|---|---|---|---|---|---| | MediaPipeFaceDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 0.559 ms | 0 - 37 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 0.632 ms | 1 - 39 MB | FP16 | NPU | [MediaPipe-Face-Detection.so](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.so) | | MediaPipeFaceDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 1.055 ms | 1 - 2 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.onnx) | | MediaPipeFaceDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.404 ms | 0 - 20 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 0.454 ms | 0 - 14 MB | FP16 | NPU | [MediaPipe-Face-Detection.so](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.so) | | MediaPipeFaceDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 0.739 ms | 0 - 42 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.onnx) | | MediaPipeFaceDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.417 ms | 0 - 12 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 0.386 ms | 0 - 12 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 0.759 ms | 0 - 31 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.onnx) | | MediaPipeFaceDetector | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 0.548 ms | 0 - 76 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 0.603 ms | 1 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | SA7255P ADP | SA7255P | TFLITE | 18.827 ms | 0 - 16 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | SA7255P ADP | SA7255P | QNN | 19.157 ms | 1 - 11 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 0.555 ms | 0 - 5 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | SA8255 (Proxy) | SA8255P Proxy | QNN | 0.607 ms | 1 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | SA8295P ADP | SA8295P | TFLITE | 1.141 ms | 0 - 10 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | SA8295P ADP | SA8295P | QNN | 1.221 ms | 0 - 6 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 0.553 ms | 0 - 5 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | SA8650 (Proxy) | SA8650P Proxy | QNN | 0.617 ms | 1 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | SA8775P ADP | SA8775P | TFLITE | 1.273 ms | 0 - 17 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | SA8775P ADP | SA8775P | QNN | 1.468 ms | 1 - 7 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 0.766 ms | 0 - 14 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.tflite) | | MediaPipeFaceDetector | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 0.827 ms | 1 - 15 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 0.748 ms | 1 - 1 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceDetector | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 1.036 ms | 2 - 2 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceDetector.onnx) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 0.196 ms | 0 - 8 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 0.278 ms | 0 - 8 MB | FP16 | NPU | [MediaPipe-Face-Detection.so](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.so) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 0.51 ms | 0 - 27 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.onnx) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.147 ms | 0 - 11 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 0.211 ms | 0 - 11 MB | FP16 | NPU | [MediaPipe-Face-Detection.so](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.so) | | MediaPipeFaceLandmarkDetector | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 0.399 ms | 0 - 32 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.onnx) | | MediaPipeFaceLandmarkDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.123 ms | 0 - 10 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 0.209 ms | 0 - 10 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 0.409 ms | 0 - 19 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.onnx) | | MediaPipeFaceLandmarkDetector | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 0.199 ms | 0 - 7 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 0.275 ms | 0 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | SA7255P ADP | SA7255P | TFLITE | 3.672 ms | 0 - 12 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | SA7255P ADP | SA7255P | QNN | 3.963 ms | 0 - 10 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 0.195 ms | 0 - 7 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | SA8255 (Proxy) | SA8255P Proxy | QNN | 0.287 ms | 0 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | SA8295P ADP | SA8295P | TFLITE | 0.576 ms | 0 - 8 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | SA8295P ADP | SA8295P | QNN | 0.791 ms | 0 - 6 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 0.191 ms | 0 - 7 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | SA8650 (Proxy) | SA8650P Proxy | QNN | 0.28 ms | 0 - 2 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | SA8775P ADP | SA8775P | TFLITE | 0.5 ms | 0 - 13 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | SA8775P ADP | SA8775P | QNN | 0.8 ms | 0 - 6 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 0.275 ms | 0 - 13 MB | FP16 | NPU | [MediaPipe-Face-Detection.tflite](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.tflite) | | MediaPipeFaceLandmarkDetector | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 0.377 ms | 0 - 16 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 0.379 ms | 0 - 0 MB | FP16 | NPU | Use Export Script | | MediaPipeFaceLandmarkDetector | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.51 ms | 3 - 3 MB | FP16 | NPU | [MediaPipe-Face-Detection.onnx](https://huggingface.co/qualcomm/MediaPipe-Face-Detection/blob/main/MediaPipeFaceLandmarkDetector.onnx) | ## Installation This model can be installed as a Python package via pip. ```bash pip install qai-hub-models ``` ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) 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. ```bash qai-hub configure --api_token API_TOKEN ``` Navigate to [docs](https://app.aihub.qualcomm.com/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. ```bash python -m qai_hub_models.models.mediapipe_face.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.mediapipe_face.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. ```bash python -m qai_hub_models.models.mediapipe_face.export ``` ``` Profiling Results ------------------------------------------------------------ MediaPipeFaceDetector Device : Samsung Galaxy S23 (13) Runtime : TFLITE Estimated inference time (ms) : 0.6 Estimated peak memory usage (MB): [0, 37] Total # Ops : 111 Compute Unit(s) : NPU (111 ops) ------------------------------------------------------------ MediaPipeFaceLandmarkDetector Device : Samsung Galaxy S23 (13) Runtime : TFLITE Estimated inference time (ms) : 0.2 Estimated peak memory usage (MB): [0, 8] Total # Ops : 100 Compute Unit(s) : NPU (100 ops) ``` ## How does this work? This [export script](https://aihub.qualcomm.com/models/mediapipe_face/qai_hub_models/models/MediaPipe-Face-Detection/export.py) leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) 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. ```python import torch import qai_hub as hub from qai_hub_models.models.mediapipe_face import Model # Load the model model = Model.from_pretrained() face_detector_model = model.face_detector face_landmark_detector_model = model.face_landmark_detector # Device device = hub.Device("Samsung Galaxy S23") # Trace model face_detector_input_shape = face_detector_model.get_input_spec() face_detector_sample_inputs = face_detector_model.sample_inputs() traced_face_detector_model = torch.jit.trace(face_detector_model, [torch.tensor(data[0]) for _, data in face_detector_sample_inputs.items()]) # Compile model on a specific device face_detector_compile_job = hub.submit_compile_job( model=traced_face_detector_model , device=device, input_specs=face_detector_model.get_input_spec(), ) # Get target model to run on-device face_detector_target_model = face_detector_compile_job.get_target_model() # Trace model face_landmark_detector_input_shape = face_landmark_detector_model.get_input_spec() face_landmark_detector_sample_inputs = face_landmark_detector_model.sample_inputs() traced_face_landmark_detector_model = torch.jit.trace(face_landmark_detector_model, [torch.tensor(data[0]) for _, data in face_landmark_detector_sample_inputs.items()]) # Compile model on a specific device face_landmark_detector_compile_job = hub.submit_compile_job( model=traced_face_landmark_detector_model , device=device, input_specs=face_landmark_detector_model.get_input_spec(), ) # Get target model to run on-device face_landmark_detector_target_model = face_landmark_detector_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. ```python face_detector_profile_job = hub.submit_profile_job( model=face_detector_target_model, device=device, ) face_landmark_detector_profile_job = hub.submit_profile_job( model=face_landmark_detector_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. ```python face_detector_input_data = face_detector_model.sample_inputs() face_detector_inference_job = hub.submit_inference_job( model=face_detector_target_model, device=device, inputs=face_detector_input_data, ) face_detector_inference_job.download_output_data() face_landmark_detector_input_data = face_landmark_detector_model.sample_inputs() face_landmark_detector_inference_job = hub.submit_inference_job( model=face_landmark_detector_target_model, device=device, inputs=face_landmark_detector_input_data, ) face_landmark_detector_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](https://myaccount.qualcomm.com/signup). ## Deploying compiled model to Android The models can be deployed using multiple runtimes: - TensorFlow Lite (`.tflite` export): [This tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a guide to deploy the .tflite model in an Android application. - QNN (`.so` export ): This [sample app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html) provides instructions on how to use the `.so` shared library in an Android application. ## View on Qualcomm® AI Hub Get more details on MediaPipe-Face-Detection's performance across various devices [here](https://aihub.qualcomm.com/models/mediapipe_face). Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/) ## License * The license for the original implementation of MediaPipe-Face-Detection can be found [here](https://github.com/zmurez/MediaPipePyTorch/blob/master/LICENSE). * The license for the compiled assets for on-device deployment can be found [here](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/Qualcomm+AI+Hub+Proprietary+License.pdf) ## References * [BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs](https://arxiv.org/abs/1907.05047) * [Source Model Implementation](https://github.com/zmurez/MediaPipePyTorch/) ## 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).