TensorRT result is different from Tensorflow results

Hi, I am running the Lanenet (https://github.com/MaybeShewill-CV/lanenet-lane-detection). I first train it on my desktop with tensorflow and python. Then I convert the tensorflow checkpoint to uff and run inference with tensorrt on Jetson Xavier. However, I found the results from tensorflow python and c++ tensort rt are slightly different. Please see the two picture below. Does anyone know why?

rear_0012-real
Tensorflow python result

rear_0012_tensorrt
TensorRT c++ result