Inference with Grayscale Image

[Rabia-Saeed]

Hi,

I have trained yolov3 (custom classes and grayscale images) and the inference seems to work good in python. I need to do inference in C++. For this reason I and using this work and the config file I have is for yolov3-tiny with channels=1. However, I have issues:

1. what(): shape '[255, 256, 1, 1]' is invalid for input of size 64992 while loading the weights. I use the same config file in python and I have no issue.
   I figures that this is related to the output layer filters and I modified them to 24 (instead of 255) as I have 3 classes. I don't have this issue anymore but a new one
2. what(): Given groups=1, weight of size [16, 3, 3, 3], expected input[1, 1, 416, 416] to have 3 channels, but got 1 channels instead
   I defined channels=1 in config during both training and inference process.
   Hence, I am unable to run it in C++.

Please let me know if you know what could be the reason. Thanks!

[Rabia-Saeed]

update: I modified previous filter in Darknet::create_modules() to 1 (instead of 3) and I don't have any errors. But my results are still not ok