--- library_name: pytorch license: other tags: - backbone - android pipeline_tag: keypoint-detection --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/web-assets/model_demo.png) # Facial-Landmark-Detection: Optimized for Qualcomm Devices Detects facial landmarks (eg, nose, mouth, etc.). This model's architecture was developed by Qualcomm. The model was trained by Qualcomm on a proprietary dataset of faces, but can be used on any image. 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/facemap_3dmm) 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/facemap_3dmm/releases/v0.60.0/facemap_3dmm-onnx-float.zip) | ONNX | w8a8 | Universal | QAIRT 2.45, ONNX Runtime 1.27.1 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/releases/v0.60.0/facemap_3dmm-onnx-w8a8.zip) | QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/releases/v0.60.0/facemap_3dmm-qnn_dlc-float.zip) | QNN_DLC | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/releases/v0.60.0/facemap_3dmm-qnn_dlc-w8a8.zip) | TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/releases/v0.60.0/facemap_3dmm-tflite-float.zip) | TFLITE | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/facemap_3dmm/releases/v0.60.0/facemap_3dmm-tflite-w8a8.zip) For more device-specific assets and performance metrics, visit **[Facial-Landmark-Detection on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/facemap_3dmm)**. ### 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/facemap_3dmm) 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 [Facial-Landmark-Detection on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.60.0/src/qai_hub_models/models/facemap_3dmm) for usage instructions. ## Model Details **Model Type:** Model_use_case.pose_estimation **Model Stats:** - Input resolution: 128x128 - Model size (float): 20.7 MB - Model size (w8a8): 5.27 MB - Number of parameters: 5.42M ## Performance Summary | Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |---|---|---|---|---|---|--- | Facial-Landmark-Detection | ONNX | float | Snapdragon® X2 Elite | 0.155 ms | 1 - 1 MB | NPU | Facial-Landmark-Detection | ONNX | float | Snapdragon® X Elite | 0.271 ms | 11 - 11 MB | NPU | Facial-Landmark-Detection | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.243 ms | 0 - 30 MB | NPU | Facial-Landmark-Detection | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 0.537 ms | 0 - 34 MB | NPU | Facial-Landmark-Detection | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.318 ms | 0 - 14 MB | NPU | Facial-Landmark-Detection | ONNX | float | Qualcomm® QCS8450 | 0.537 ms | 0 - 34 MB | NPU | Facial-Landmark-Detection | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 0.44 ms | 0 - 3 MB | NPU | Facial-Landmark-Detection | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 0.271 ms | 11 - 11 MB | NPU | Facial-Landmark-Detection | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 0.216 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | ONNX | float | Snapdragon® 8 Elite Mobile | 0.216 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.213 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® X2 Elite | 0.074 ms | 1 - 1 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® X Elite | 0.124 ms | 5 - 5 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.138 ms | 0 - 39 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.267 ms | 0 - 45 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 0.728 ms | 0 - 3 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.362 ms | 0 - 3 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.169 ms | 0 - 8 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® QCS8450 | 0.267 ms | 0 - 45 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.263 ms | 0 - 3 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 0.124 ms | 5 - 5 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 0.616 ms | 0 - 26 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 0.241 ms | 0 - 26 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.12 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® 8 Elite Mobile | 0.12 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.119 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.241 ms | 0 - 26 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® X2 Elite | 0.233 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® X Elite | 0.369 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.232 ms | 0 - 29 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 0.545 ms | 0 - 34 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 0.422 ms | 0 - 3 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 1.167 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.309 ms | 0 - 26 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® SA8775P | 0.503 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® SA8650P | 0.503 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® SA8255P | 0.503 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® QCS8450 | 0.545 ms | 0 - 34 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 0.391 ms | 2 - 4 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 0.369 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 0.215 ms | 0 - 25 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® SA7255P | 1.167 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Qualcomm® SA8295P | 0.709 ms | 0 - 17 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® 8 Elite Mobile | 0.215 ms | 0 - 25 MB | NPU | Facial-Landmark-Detection | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.202 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.17 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® X Elite | 0.223 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.137 ms | 0 - 37 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.237 ms | 0 - 42 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 0.752 ms | 0 - 2 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.221 ms | 0 - 2 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.447 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.169 ms | 0 - 60 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® SA8775P | 0.309 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® SA8650P | 0.309 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® SA8255P | 0.309 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 0.237 ms | 0 - 42 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.224 ms | 2 - 4 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 0.223 ms | 0 - 0 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 0.595 ms | 0 - 24 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 0.218 ms | 0 - 24 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.114 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® SA7255P | 0.447 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Qualcomm® SA8295P | 0.445 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® 8 Elite Mobile | 0.114 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.116 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.218 ms | 0 - 24 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.222 ms | 0 - 40 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 0.479 ms | 0 - 46 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 0.411 ms | 0 - 13 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 1.141 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.277 ms | 0 - 37 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® SA8775P | 0.487 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® SA8650P | 0.487 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® SA8255P | 0.487 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® QCS8450 | 0.479 ms | 0 - 46 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 0.376 ms | 0 - 13 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 0.203 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® SA7255P | 1.141 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Qualcomm® SA8295P | 0.682 ms | 0 - 18 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Snapdragon® 8 Elite Mobile | 0.203 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.195 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.133 ms | 0 - 36 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.234 ms | 0 - 42 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 0.658 ms | 0 - 7 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.239 ms | 0 - 8 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.474 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.211 ms | 0 - 2 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® SA8775P | 0.334 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® SA8650P | 0.334 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® SA8255P | 0.334 ms | 0 - 22 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® QCS8450 | 0.234 ms | 0 - 42 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.226 ms | 0 - 7 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 0.603 ms | 0 - 24 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 0.217 ms | 0 - 23 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.124 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® SA7255P | 0.474 ms | 0 - 21 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Qualcomm® SA8295P | 0.475 ms | 0 - 18 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Snapdragon® 8 Elite Mobile | 0.124 ms | 0 - 19 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.118 ms | 0 - 20 MB | NPU | Facial-Landmark-Detection | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.217 ms | 0 - 23 MB | NPU ## License * The license for the original implementation of Facial-Landmark-Detection can be found [here](https://github.com/qualcomm/ai-hub-models/blob/main/LICENSE). ## 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).