Unsupervised Domain Adaptation for Semantic Segmentation using Depth Distribution

Authors: Quanliang Wu, Huajun Liu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on SYNTHIA-to-Cityscapes and SYNTHIAto-Mapillary benchmarks show the effectiveness of the method.
Researcher Affiliation Academia Quanliang Wu, Huajun Liu School of Computer Science Wuhan University, Wuhan, China {quanliangwu, huajunliu}@whu.edu.cn
Pseudocode Yes Algorithm 1: Spatial prior pseudo-labels refinement algorithm
Open Source Code Yes The source code is available at https://github.com/depdis/Depth_Distribution.
Open Datasets Yes For the SYNTHIA dataset [7] is used as the source domain. Following [12, 13], we use the SYNTHIA-RAND-CITYSCAPES split consisting of 9,400 synthetic images and their corresponding pixel-wise semantic labels and depth. For target domains, we use Cityscapes [8] and Mapillary Vistas [9] datasets.
Dataset Splits Yes Following [12, 13], we use the SYNTHIA-RAND-CITYSCAPES split consisting of 9,400 synthetic images and their corresponding pixel-wise semantic labels and depth. Similar to [26, 19, 12, 13], we report the performance of semantic segmentation based on mean Intersection over Union (m Io U in %) on the 16 classes of the Cityscapes validation set
Hardware Specification Yes All our experiments are conducted on a single NVIDIA 1080Ti GPU with a memory of 11GB.
Software Dependencies No The paper mentions "Our network is implemented on Py Torch [40]" but does not specify its version number or versions for other key software dependencies.
Experiment Setup Yes The learning rates of the prediction and discriminator networks are set as 2.5 10 4 and 1.0 10 3 respectively. In self-training, the parameters are: Q1 = 54K, Q2 = 30K. In our experiments, we use λseg = 1.0, λdep =0.5 10 2, λbal = 10 2, λtar = 5 10 2.