Architecture | Documentation | Blog | Roadmap | Slack
🌐 Language | 语言： 🇨🇳 中文
This repository primarily provides inference capabilities for multi-task networks in both 2D and 3D. It includes packaged libraries to support daily development, integration, testing, and inference. The framework implements multithreading, the singleton pattern, and producer-consumer patterns. It also supports cache log analysis.
| Libraries | Eigen | Gflags | Glog | Yaml-cpp | Cuda | Cudnn | Tensorrt | Opencv |
|---|---|---|---|---|---|---|---|---|
| Version | 3.4 | 2.2.2 | 0.6.0 | 0.8.0 | 11.4 | 8.4 | 8.4 | 3.4.5 |
Visit our documentation to learn more.
 |
 |
|---|---|
 |
 |
- Dataset:
- BDD100K
The validation dataset is BDD100K, which contains 70000 training samples and 10000 val samples. All models in the table were trained on the full BDD100K dataset.
- nuscenes
The validation dataset is nuscenes-mini. All models in the table were trained on the full nuscenes dataset.
- BDD100K
- Model: The deployed model is the 's' version of the YOLO multi-task network series.
- Quantize: Quantization was performed using NVIDIA's Post-Training Quantization (PTQ) method.
| Model | Platform | Resolution | mAP50-95(fp32) | mAP50(fp32) | mAP50-95(fp16) | mAP50(fp16) | mAP50-95(int8) | mAP50(int8) | fps(fp32) |
|---|---|---|---|---|---|---|---|---|---|
| A-YOLOM | RTX4060 | 480x640 | - | - | - | - | - | - | 61.8229 |
| Orin x | 480x640 | - | - | - | - | - | - | - | |
| Thor | 480x640 | - | - | - | - | - | - | - |
Welcome users to participate in these projects. Please refer to CONTRIBUTING.md for the contributing guideline.We encourage you to join the effort and contribute feedback, ideas, and code. You can participate in Working Groups, Working Groups have most of their discussions on Slack or QQ (938558640).
- Add YOLOP model
- Add Thor platform support
- Add quantization support to the model
- Add API support for versions greater than TRT 8.0