Darknet: No detection of Tiny yolo on raspberry pi 3 with custom weights of custom dataset

[abhigoku10]

@pjreddie @tjluyao Hi guys , i trained my weights for custom dataset i am having a strange issues when i test the weights on PI3 no detection is happening but on cpu/gpu machines i get the detection with bounding boxes using the same weights. The strange thing is when i use the tiny-yolo.weights which is default i am able to see the bounding box on an image , can you suggest me any alternative for it . The following are the trial experiments i have done
1 .Changed the threshold

1. Re-trained and used to the new weights and tested
   3.Tested with default yolo weights ( i get the probability and bounding boxes)

Thanks in advance

[steph1793]

Hi @abhigoku10 ! I had exactly the same issue. I found the solution here. In fact, in the file src/parser.c, sometimes we need sizeof(size_t). But sizeof(size_t) on a 64bit computer equals 8, while it worths 4 on a raspberry (weights you have are probably trained on a 64 bit computer). So you must edite this file (src/parser.c) on your raspberry, by replacing all the sizeof(size_t) by [sizeof(size_t)*2]) . Then, run make to regenerate the darknet executable file.

[alapha23]

@steph1793
Thank you so much for the answer!
I had quite a hard time and your reply saved my model.
Perhaps they should implement this in their makefile to support darknet on Rasberry pi.