--- library_name: pytorch license: other tags: - bu_auto - real_time - android pipeline_tag: gaze-estimation --- ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/eyegaze/web-assets/model_demo.png) # EyeGaze: Optimized for Qualcomm Devices Predicts gaze direction (pitch, yaw) from 96x160 grayscale eye images using the EyeNet model. This is based on the implementation of EyeGaze found [here](https://github.com/david-wb/gaze-estimation). 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/main/src/qai_hub_models/models/eyegaze) 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 | ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/eyegaze/releases/v0.51.0/eyegaze-onnx-float.zip) | ONNX | w8a16 | Universal | ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/eyegaze/releases/v0.51.0/eyegaze-onnx-w8a16.zip) | QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/eyegaze/releases/v0.51.0/eyegaze-qnn_dlc-float.zip) | TFLITE | float | Universal | | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/eyegaze/releases/v0.51.0/eyegaze-tflite-float.zip) For more device-specific assets and performance metrics, visit **[EyeGaze on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/eyegaze)**. ### Option 2: Export with Custom Configurations Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/eyegaze) 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 [EyeGaze on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/eyegaze) for usage instructions. ## Model Details **Model Type:** Model_use_case.gaze_estimation **Model Stats:** - Model checkpoint: checkpoint.pt - Input resolution: 96x160 - Number of parameters: 2.58M - Model size (float): 9.6MB - Model size (w8a16): 3.3 MB ## Performance Summary | Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |---|---|---|---|---|---|--- | EyeGaze | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 19.064 ms | 32 - 43 MB | CPU | EyeGaze | ONNX | float | Snapdragon® X2 Elite | 9.205 ms | 35 - 35 MB | CPU | EyeGaze | ONNX | float | Snapdragon® X Elite | 9.936 ms | 34 - 34 MB | CPU | EyeGaze | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 22.114 ms | 31 - 41 MB | CPU | EyeGaze | ONNX | float | Qualcomm® QCS8550 (Proxy) | 25.588 ms | 27 - 48 MB | CPU | EyeGaze | ONNX | float | Qualcomm® QCS9075 | 20.894 ms | 33 - 43 MB | CPU | EyeGaze | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 19.938 ms | 32 - 44 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 25.828 ms | 69 - 84 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® X2 Elite | 17.489 ms | 102 - 102 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® X Elite | 19.36 ms | 101 - 101 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 33.106 ms | 71 - 83 MB | CPU | EyeGaze | ONNX | w8a16 | Qualcomm® QCS6490 | 143.117 ms | 69 - 74 MB | CPU | EyeGaze | ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 38.98 ms | 66 - 75 MB | CPU | EyeGaze | ONNX | w8a16 | Qualcomm® QCS9075 | 40.471 ms | 68 - 73 MB | CPU | EyeGaze | ONNX | w8a16 | Qualcomm® QCM6690 | 68.089 ms | 71 - 82 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 27.141 ms | 70 - 82 MB | CPU | EyeGaze | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 59.86 ms | 85 - 96 MB | CPU | EyeGaze | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 12.849 ms | 10 - 25 MB | CPU | EyeGaze | QNN_DLC | float | Snapdragon® X2 Elite | 9.742 ms | 11 - 11 MB | CPU | EyeGaze | QNN_DLC | float | Snapdragon® X Elite | 13.518 ms | 11 - 11 MB | CPU | EyeGaze | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 17.404 ms | 10 - 23 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 214.415 ms | 10 - 24 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 20.21 ms | 10 - 64 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® SA8775P | 28.528 ms | 11 - 21 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® QCS9075 | 67.542 ms | 76 - 150 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 26.795 ms | 11 - 25 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® SA7255P | 214.415 ms | 10 - 24 MB | CPU | EyeGaze | QNN_DLC | float | Qualcomm® SA8295P | 23.19 ms | 10 - 20 MB | CPU | EyeGaze | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 14.301 ms | 10 - 26 MB | CPU | EyeGaze | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 10.804 ms | 2 - 12 MB | CPU | EyeGaze | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 16.076 ms | 2 - 12 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 30.304 ms | 2 - 10 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 18.487 ms | 2 - 51 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® SA8775P | 25.293 ms | 4 - 9 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® QCS9075 | 25.271 ms | 2 - 43 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 21.712 ms | 2 - 12 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® SA7255P | 30.304 ms | 2 - 10 MB | CPU | EyeGaze | TFLITE | float | Qualcomm® SA8295P | 16.681 ms | 2 - 7 MB | CPU | EyeGaze | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 11.561 ms | 2 - 9 MB | CPU ## License * The license for the original implementation of EyeGaze can be found [here](https://github.com/quic/ai-hub-models/blob/main/LICENSE). ## References * [Source Model Implementation](https://github.com/david-wb/gaze-estimation) ## 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).