GTA2CS running on 16 classes

[Masrur02]

Hi, it seems GTA2CS is getting validated on 16 classes. Train, Bicycle, Rider class shows 0 mIoU.

+---------------+-------+-------+
| Class | IoU | Acc |
+---------------+-------+-------+
| road | 91.41 | 94.06 |
| sidewalk | 54.69 | 78.73 |
| building | 87.08 | 93.17 |
| wall | 49.62 | 59.68 |
| fence | 33.68 | 39.25 |
| pole | 49.17 | 64.76 |
| traffic light | 49.72 | 58.69 |
| traffic sign | 21.12 | 21.31 |
| vegetation | 88.88 | 95.15 |
| terrain | 50.7 | 67.45 |
| sky | 87.18 | 99.42 |
| person | 60.74 | 90.28 |
| rider | 0.0 | 0.0 |
| car | 85.77 | 97.08 |
| truck | 59.02 | 85.02 |
| bus | 47.7 | 66.83 |
| train | 0.0 | 0.0 |
| motorcycle | 2.8 | 2.8 |
| bicycle | 0.0 | 0.0 |
+---------------+-------+-------+
2025-05-30 08:11:08,316 - mmseg - INFO - Summary:
2025-05-30 08:11:08,316 - mmseg - INFO -
+-------+-------+-------+
| aAcc | mIoU | mAcc |
+-------+-------+-------+
| 90.46 | 48.38 | 58.62 |

[cfmata]

Hi @Masrur02,

Apologies for not catching this earlier. I'm assuming you've trained a model on GTA->CS and would like to check mIOU on CS val. How are you running the evaluation? You can adapt this test command from MIC and run from the CoPT root:

python -m tools.test configs/mic/gta2cs_copt.py --eval mIoU

Let me know if that helps.

[Masrur02]

Hi, before using the evaluation python -m tools.test configs/mic/gta2cs_copt.py --eval mIoU, I was checking the result during training after each 4000 iterations, and that gave me 0 mIoU in 3 classes.

[cfmata]

Can you confirm whether this happens throughout the entire training and not just early on? My rider class starts to get non-zero IOU after 4k iterations.

[Masrur02]

I did not try it yet. However I saw another issue, it seems you are getting less IoU for the class train in MIC setup. Do you know why so? And also how did you overcome that in your CoPT