Luckfox Forums

Hello everyone,

I am currently working on a project using the Luckfox Pico Ultra W RV1106 board. My goal is to build a real-time smoke and fire detection system.

I have been researching the best approach for this and have a few questions regarding model compatibility and performance on this specific board.

1. Supported Models: I've seen that models like YOLOv5 and RetinaFace are officially mentioned in the documentation. Are there other object detection models, such as YOLOv8 or others, that are known to work well on the RV1106 NPU? I'm looking for recommendations for a model that provides a good balance of accuracy and performance for a smoke/fire detection task.
2. YOLOv5 Version Limits: If I decide to use YOLOv5, I understand that the smaller versions like YOLOv5n and YOLOv5s are recommended due to the hardware constraints of the RV1106. However, I would like to confirm if it's feasible to run larger versions like YOLOv5m, YOLOv5l, or YOLOv5x. Has anyone successfully deployed these larger models, and what was the performance like? I'm concerned about memory usage and inference speed.

Any advice, performance benchmarks, or shared experiences from the community would be greatly appreciated.

Thank you in advance

Hello, we previously learned that some users were running YOLOV10 on RV1106, but the users did not share the usage methods and performance results.
The performance of the model is not the only factor determining whether RV1106 can be used. The most important thing is to ensure that all the operators required by the model, RV1106's NPU, are supported. If the operators in the model you are using are incompatible, you will need to modify the model or add operators. This is not an easy task and requires a good understanding of the model and RKNN operator registration. It is recommended that you start with rockchip's default-supported yolov5s.