Results

we claim that we utilise ONE single model to perform all the evaluations to simulate driving scenarios in real life.

cityscapes reports

methods	mIoU (class)
closed seg. [paper] [code]	82.46 [report]
dense-hybrid [paper] [code]	82.06 [report]
meta-ood [paper] [code]	81.51 [report]
pebal [paper] [code]	81.19 [report]
ours	82.46 [report]

checkpoints

you can reproduce our results based on the supported checkpoints below:

• rpl can download in here, and rpl+corocl can download in here.

for the segment-me-if-you-can (SMIYC), download the official evaluation code (with an extra post-process stage) to achieve the reported performance. we support the prediction & results in here.

training details

1. you can download our training log via this LINK.
2. for more details, you can check our wandb log in this LINK, where it includes:
   1. overall information (e.g., training command line, hardware information and training time).
   2. training details (e.g., loss curves, validation results and visualization)
   3. output logs (well, sometimes might crash ...)

• in the final training stage, we adopt longer training epochs and more frequent validation to choose the potential best model for black boxing test sets.