Why is my reproduction result higher than the one in the paper?

[torotocat]

1. class：| person | rider | car | truck | bus | train | motorcycle | bicycle |
2. my: | 58.419 | 60.035 | 70.148 | 45.800 | 62.820 | 47.403 | 48.602 | 59.398 |
3. paper: |52.1 |55.8 |67.5 | 32.7 | 55.9 | 49.1 | 40.1 | 50.3 |

What is the cause of this issue? My Cityscapes and Foggy datasets were converted to VOC format using your script. For experimental rigor, I also used a different script to generate the VOC dataset, but the result is the same. Additionally, is the beta index for the Foggy dataset simply 0.02, or is there something else?

When reproducing the results, the only changes I made were updating a few packages due to conflicts; the rest of the environment setup is the same as in the paper.

[kinredon]

Yes, the foggy level is 0.02. Although the code in this repo may produce better results such as ~51.0, your results are abnormal and much higher than expectation. Based on the information provided by you, I cannot figure out what reason causing this issue. Do you directly train the model on the foggy cityscapes with labeled data?