Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters

Authors: Dong Zhao, Qi Zang, Shuang Wang, Nicu Sebe, Zhun Zhong

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that Se Co can be flexibly applied to various cross-domain semantic segmentation tasks, i.e. domain generalization and domain adaptation, even including source-free, and black-box domain adaptation, significantly improving the performance of existing state-of-the-art methods.
Researcher Affiliation Academia Dong Zhao1 , Qi Zang1 , Shuang Wang1 , Nicu Sebe2, Zhun Zhong3,4 1 School of Artificial Intelligence, Xidian University, Shaanxi, China 2 Department of Information Engineering and Computer Science, University of Trento, Italy 3 School of Computer Science and Information Engineering, Hefei University of Technology, China 4 School of Computer Science, University of Nottingham, NG8 1BB Nottingham, UK
Pseudocode Yes Algorithm 1 Aggregation of Pseudo-Labels with SAM
Open Source Code Yes The code is available at https://github.com/DZhao Xd/Se Co.
Open Datasets Yes Datasets. We employ two real datasets (Cityscapes [12] and BDD-100k [86]) alongside two synthetic datasets (GTA5 [63] and SYNTHIA [65]). The details of these datasets are introduced in Section B.
Dataset Splits Yes The Cityscapes dataset comprises 2,975 training images and 500 validation images, all with a resolution of 2048 1024.
Hardware Specification No The paper mentions using 'SAM [36] with Vision Transformer-H (Vi T-H) [14]' but does not specify the hardware (e.g., GPU models, CPU, RAM) used for running the experiments or training the models.
Software Dependencies No The paper mentions using Deep Lab V2, Seg Former, and SAM, but does not specify software versions for libraries, frameworks, or operating systems (e.g., PyTorch 1.x, CUDA 11.x).
Experiment Setup Yes In Algorithm 1, the enlargement factor for the bounding box area is set to 1.5. The connectivity classifier is trained only for 5000 iterations in an early learning way for all tasks. The noise threshold (τns) and correction threshold (τcr) are configured at 0.60 and 0.95, respectively.