notesum.ai
Published at November 25A Study on Unsupervised Domain Adaptation for Semantic Segmentation in the Era of Vision-Language Models
cs.CV
cs.AI
Released Date: November 25, 2024
Authors: Manuel Schwonberg1, Claus Werner2, Hanno Gottschalk1, Carsten Meyer3
Aff.: 1TU Berlin; 2CARIAD SE; 3Ostfalia University of Applied Sciences
| GTA5Cityscapes | |||||||||||
| Architecture | UDA Method (in % mIoU) | ||||||||||
| Encoder | Decoder | AdaptSegNet [Tsai et al.(2018)Tsai, Hung, Schulter, Sohn, Yang, and Chandraker] | ADVENT [Vu et al.(2019)Vu, Jain, Bucher, Cord, and Perez] | DACS [Tranheden et al.(2021)Tranheden, Olsson, Pinto, and Svensson] | DAFormer [Hoyer et al.(2022a)Hoyer, Dai, and Van Gool] | Src. Only | Oracle | ||||
| Abs. | Rel. | Abs. | Rel. | Abs. | Rel. | Abs. | Rel. | ||||
| ResNet-101 | DeepLabV2 | 44.0 | 60.4 | 43.4 | 59.6 | 55.3 | 76.0 | 57.9 | 79.5 | 36.0 | 72.8 |
| MiT-B5 | DAFormer | 47.5 | 60.8 | 47.1 | 60.3 | 60.3 | 77.2 | 67.6 | 86.6 | 47.6 | 78.1 |
| EVA02-L | DAFormer | 59.5 | 73.1 | 59.2 | 72.7 | 70.3 | 86.4 | 68.0 | 83.5 | 60.5 | 81.4 |