Weather Change

EXPERIMENT 3
From a clear sunset to a hard rain scenario

It tests the generalization and adaptation performances of the semantic segmentation networks when changing the weather condition, from a sunset to a hard rain scene. The gap among the source and target domain is high but lower than the shift across two different towns. The difficulty of the task is medium.

Source Scenario:

  • Environment: Town 01
  • Weather and illumination condition: Clear Sunset
  • Viewpoint: Jeep Wrangler Rubicon
  • Source train set size: 10044
Source Example Source Example Source Example Source Example Source Example Source Example
Click to see some sample taken from the source scenario.

Target/Test Scenario:

  • Environment: Town 01
  • Weather and illumination condition: Hard Rain Noon
  • Viewpoint: Jeep Wrangler Rubicon
  • Target train set size (only with DA): 10008
  • Test size: 1668
Target Example Target Example Target Example Target Example Target Example Target Example
Click to see some sample taken from the target/test scenario.

Results

Experiment Distance Measurement Performance Evaluation
Euclidean
distance		 Cosine
distance		 Bhattacharaya
distance		 Network Code Available mIoU (%)
Source:
Town 01, Clear Sunset, Jeep

Target:
Town 01, Hard Rain Noon, Jeep		 5,6555 1,2633 0,0337 without
domain
adaptation 		DeepLab V2[1] (soon) 40,24
DeepLab V3+ [2] (soon) 33,93
PSPNet [3] (soon) 29,65
PSANet [4] (soon) 33,60
DeepLab V2 [1]
(source=target)		 (soon) 78,31
with
domain
adaptation		 DADA [5] (soon) 55,87
ADVENT [6] (soon) 61,13
CLAN [7] (soon) 65,52
DISE [8] (soon) 71,91

Updated: