Page 1 of 3
Help required for running a custom object detection model into pico pi board
Posted: 2025-08-29 5:21
by Infitechlab1
Hello everyone,
I have luckfox pico pi board, on which I am trying to run a object detection AI. I went through and tried your examples given on the github and wiki. But what I dident found was how to deploy a custom model on the board.
I have some small models from edge impulse and localy trained model which I would like to run and test on the pico pi board. Kindly guide me throught this.
Thank you
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-03 11:03
by Infitechlab1
Kindly, someone respond. Luckfox team.
As I am reading 1106 has limited support for AI. Kindly guide me through this, how can a custom model be deploied into the board.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-05 3:08
by Crocodile
Hello, the deployment of custom models involves the compatibility of convolutional neural network operators and the implementation of post-processing solutions. Different models require different solutions, and it is necessary to have a thorough understanding of the model to complete the migration work.
Therefore, we do not provide technical support for the migration of non-officially supported models. We can only recommend that you learn the migration methods from the wiki and apply them to other migration tasks.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-09 6:41
by Infitechlab1
Hi, as per the wiki and luckfox github, we did all the process of creating our very own model and then converted it too rknn and ran it. All is good but the issue is the file format you have given, lets say for yolov5 has a class file(.txt), and onnx or rknn does not require lable file if the classes are hard coded in the code, one issue is that the file sequence or usage of file.
Second is there is "anchors_yolov5.txt", how it has been created we are not getting, cause while converting to ONNX model it was not created, then even when the rknn folder was created it was not there, and if we are not wrong we are getting a "segmentantion fault", cause we just replaced the required file in your yolov5 example, and executed it.
See your wiki and all the document are very detailed but are scattered accross multiple places, what we expect is a very basic yet proper AI file which we can try and understand how the AI runs on RKNPU. Cause how to train the model is available on Ultralytics github, but the proper way of arranging the model file and what files are required is not mentioned, for which we are relying on your docs.
Kindly guide us.
Thank you.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-12 7:32
by Infitechlab1
Hi,
Can someone please tell me how to run a custom yolo model in luckfox 1103 or 1106 board? I have been going through the process again and again as per the luckfox wiki and github, but no output. The process given by luckfox dosent mention anything about custom models. Kindly guide me, luckfox team, anyone who has sucessfull ran a custom yolov5 model on the pico pi, pico pi ultra, mini B.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-12 15:18
by pointer
Infitechlab1 wrote: ↑2025-09-12 7:32
Hi,
Can someone please tell me how to run a custom yolo model in luckfox 1103 or 1106 board? I have been going through the process again and again as per the luckfox wiki and github, but no output. The process given by luckfox dosent mention anything about custom models. Kindly guide me, luckfox team, anyone who has sucessfull ran a custom yolov5 model on the pico pi, pico pi ultra, mini B.
Actually, you can follow a standard YOLOv5 training tutorial, with the only differences being that when cloning the repository, you should replace
https://github.com/ultralytics/yolov5 with
https://github.com/airockchip/yolov5, and later when converting the .pt model to .onnx using export.py, you need to add the --rknpu flag; as for converting .onnx to .rknn, you can directly follow the instructions mentioned in the Luckfox Wiki.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-12 15:34
by pointer
Infitechlab1 wrote: ↑2025-09-09 6:41
Hi, as per the wiki and luckfox github, we did all the process of creating our very own model and then converted it too rknn and ran it. All is good but the issue is the file format you have given, lets say for yolov5 has a class file(.txt), and onnx or rknn does not require lable file if the classes are hard coded in the code, one issue is that the file sequence or usage of file.
Second is there is "anchors_yolov5.txt", how it has been created we are not getting, cause while converting to ONNX model it was not created, then even when the rknn folder was created it was not there, and if we are not wrong we are getting a "segmentantion fault", cause we just replaced the required file in your yolov5 example, and executed it.
See your wiki and all the document are very detailed but are scattered accross multiple places, what we expect is a very basic yet proper AI file which we can try and understand how the AI runs on RKNPU. Cause how to train the model is available on Ultralytics github, but the proper way of arranging the model file and what files are required is not mentioned, for which we are relying on your docs.
Kindly guide us.
Thank you.
In my usage, I didn’t utilize the anchors_yolov5.txt file; this is a modified demo I created. You can specify the model and label paths using the -m and -l flags, which provides a convenient way to verify whether the YOLOv5 RKNN model is valid. However, note that the label classes are automatically determined based on the line count in the label file. If this doesn’t align with the model’s classes, it will cause the program to hang. Therefore, please avoid adding extra blank lines. For additional usage instructions, you can refer to the --help option.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-15 12:54
by Infitechlab1
Hi @pointer, see after doing all the process again, I have converted the model to rknn, but now i am not getting how to run it on device without adb. Many of the things are still not clear, how to run the ai on board after converting the model to rknn format. If you want to have a my rknn mode, will attach it. Kindly have a look and let me know.
Thank you.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-22 4:08
by Infitechlab1
Hello,
Kindly revert me back on my issue. I have given you my model too, for validation and testing perposes. I have been following all the steps directed by you. Kindly help me with very simple problem of how to run a custom model on luckfox device's, in my case pico pi w or any other 1106/1103 based device manufactured by luckfox. I have most of them. I have been trying to get this sorted for long now.
Thank you.
Re: Help required for running a custom object detection model into pico pi board
Posted: 2025-09-23 9:49
by Infitechlab1
Infitechlab1 wrote: ↑2025-09-15 12:54
Hi @pointer, see after doing all the process again, I have converted the model to rknn, but now i am not getting how to run it on device without adb. Many of the things are still not clear, how to run the ai on board after converting the model to rknn format. If you want to have a my rknn mode, will attach it. Kindly have a look and let me know.
Thank you.
Hi,
I have attached my rknn model in this post, if anyone from luckfor or any other user has ever runned a custom yolov5 model in the 1106/1103 based board, kindlt let me know. I am still trying what is possible. Kindly help.
Thank you.