# No code snippets available yet for this library.
# To use this model, check the repository files and the library's documentation.
# Want to help? PRs adding snippets are welcome at:
# https://github.com/huggingface/huggingface.jsAIPV1EMI0001 — YOLOv7-tiny Object Detector for EdgeAI Station
Detect cars, people and trucks on an EdgeAI Station by EDGEMATRIX with YOLOv7-tiny. This repository provides a 640 x 640 ONNX model and a prebuilt FP16 TensorRT engine for deployment with NVIDIA DeepStream and Triton.
Get started with the step-by-step build notebook to build and package the AI Processor on your station. The companion GitHub repository also includes the inference configuration, Node-RED flows and a live detection dashboard.
Files
| File | Description |
|---|---|
yolov7tiny-640.onnx |
Primary artifact. ONNX export with EfficientNMS_TRT, consumed by the repository's notebook and bin/build.sh. |
yolov7-fp16-1x8x8-jp6-l4t36.3-arm64.engine |
Prebuilt TensorRT FP16 engine, dynamic batch 1-8, for the validated Jetson Orin NX / JetPack 6.0 / L4T R36.3 / TensorRT 8.6.2 environment. Rebuild from ONNX for other targets and validate compatibility. |
build_info.json |
Platform details, batch profile, artifact checksums and a rebuild command |
SHA256SUMS |
Checksums for artifact integrity verification |
LICENSE |
GPL-3.0 licence text for the model artifacts |
Model details
- Architecture: YOLOv7-tiny
- Classes:
car,person,truck - Input: NCHW layout, 640 x 640 pixels
- Export: ONNX with
EfficientNMS_TRTpost-processing - TensorRT engine: FP16, dynamic batch size 1-8
- Validated target: Jetson Orin NX 16GB, JetPack 6.0, L4T R36.3, TensorRT 8.6.2
The original training dataset and exact upstream checkpoint are not documented. No accuracy evaluation is provided with these artifacts.
Performance
Observed with trtexec on a Jetson Orin NX 16GB (EdgeAI Station, FP16, 640 x 640)
while another pipeline was running: approximately 99 images/s at batch 1,
101 images/s at batch 4 and 104 images/s at batch 8.
These are shared-station measurements, not guaranteed performance or end-to-end video rates.
A separately built batch-4 engine was within 5% of the batch 1-8 engine in this comparison;
only the batch 1-8 engine is published.
Usage
Follow the Jupyter notebook in the GitHub repository. It downloads the selected artifact and verifies its pinned SHA-256, builds the custom parser and optionally the TensorRT engine, validates and exports the AIP package, and provides pipeline and Node-RED integration steps.
Attribution & License
The model architecture is YOLOv7 by WongKinYiu et al.:
Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao. "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors." CVPR 2023. https://github.com/WongKinYiu/yolov7 — GPL-3.0 License.
The ONNX model and prebuilt TensorRT engine are published under GPL-3.0;
see LICENSE. EdgeMatrix provides the EdgeAI Station integration and
deployment packaging.
The AI Processor source code in edge-ai/AIPV1EMI0001 is licensed separately under Apache-2.0, with the third-party notices described in that repository.
- Downloads last month
- 43