Efficient and Robust 2D-to-BEV Representation Learning via Geometry-guided Kernel Transformer
Paper • 2206.04584 • Published
GKT: Optimized for Qualcomm Devices
Geometry-guided Kernel Transformer is a machine learning model for generating a birds eye view represenation from the sensors(cameras) mounted on a vehicle.
This is based on the implementation of GKT 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 |
|---|---|---|---|---|
| PRECOMPILED_QNN_ONNX | float | Snapdragon® X2 Elite | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 1 Mobile | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Mobile | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.45, ONNX Runtime 1.27.1 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® X2 Elite | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 1 Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-8275 | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-9075 | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Mobile | QAIRT 2.45 | Download |
| QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | QAIRT 2.45 | Download |
For more device-specific assets and performance metrics, visit GKT 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 GKT on GitHub for usage instructions.
Model Details
Model Type: Model_use_case.driver_assistance
Model Stats:
- Input resolution: 1 x 6 x 3 x 224 x 480
- Model checkpoint: map_segmentation_gkt_7x1_conv_setting2.ckpt
- Model size: 4.66 MB
- Number of parameters: 1.18M
Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit |
|---|---|---|---|---|---|---|
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® X2 Elite | 57.794 ms | 8 - 8 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® X Elite | 103.379 ms | 7 - 7 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 3 Mobile | 82.81 ms | 10 - 18 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Gen 1 Mobile | 210.778 ms | 8 - 21 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 102.663 ms | 7 - 11 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 113.996 ms | 8 - 13 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® QCS8450 | 210.778 ms | 8 - 21 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 112.086 ms | 7 - 10 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 103.379 ms | 7 - 7 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 71.758 ms | 1 - 8 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Mobile | 71.758 ms | 1 - 8 MB | NPU |
| GKT | PRECOMPILED_QNN_ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 59.217 ms | 2 - 9 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® X2 Elite | 61.888 ms | 7 - 7 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® X Elite | 104.653 ms | 7 - 7 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 3 Mobile | 82.114 ms | 8 - 15 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Gen 1 Mobile | 199.12 ms | 8 - 20 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-8275 | 102.138 ms | 7 - 17 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-8275 | 190.133 ms | 0 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 109.338 ms | 8 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® SA8775P | 112.636 ms | 0 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® SA8650P | 112.636 ms | 0 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® SA8255P | 112.636 ms | 0 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® QCS8450 | 199.12 ms | 8 - 20 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-9075 | 110.854 ms | 7 - 17 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ IQ-X7181 | 104.653 ms | 7 - 7 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® Dragonwing™ Q-8750 | 69.951 ms | 7 - 16 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® SA7255P | 190.133 ms | 0 - 10 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Qualcomm® SA8295P | 142.388 ms | 0 - 6 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Mobile | 69.951 ms | 7 - 16 MB | NPU |
| GKT | QNN_CONTEXT_BINARY | float | Snapdragon® 8 Elite Gen 5 Mobile | 58.601 ms | 7 - 16 MB | NPU |
License
- The license for the original implementation of GKT can be found here.
References
- Efficient and Robust 2D-to-BEV Representation Learning via Geometry-guided Kernel Transformer
- 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.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support