Hi @LChachayma, yes if you reduce the number of object classes, the performance will increase.
I am currently working on an object detection transfer learning tutorial that uses PyTorch and ONNX with SSD-Mobilenet. The current SSD-Mobilenet/Inception models were trained with TensorFlow, and it can be a bit complicated to convert the models. However you can see some recommendations of existing TensorFlow object detection tutorials here: https://forums.developer.nvidia.com/t/digits-or-somthing-else/110790/7?u=dusty_nv