search

lhoyer / HRDA

[ECCV22] Official Implementation of HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation
Other
231 stars 32 forks source link

Questions on Sliding Window Inference #18

Closed JoinWorldMC closed 1 year ago

JoinWorldMC commented 1 year ago

In the appendix, 'whole' [2] does not use sliding window inference. So, what does [1] mean?

[1] hr_slide_inference = False [2] test_cfg=dict(mode='whole')

lhoyer commented 1 year ago

HRDA typically uses two levels of cropping during training. The outer crop is used to capture LR context while the inner crop is used to capture HR details. [2] means that the whole image is used as the outer crop during inference, while [1] means that the whole outer crop is used as the inner crop during inference.