Open Compound Domain Adaptation with Object Style Compensation for Semantic Segmentation

Authors: Tingliang Feng, Hao Shi, Xueyang Liu, Wei Feng, Liang Wan, Yanlin Zhou, Di Lin

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an intensive evaluation of the object style compensation. On the public datasets (e.g., C-Driving [7], ACDC [12], Cityscapes [13], KITTI [14], and Wild Dash [15]) that allows OCDA to assist semantic segmentation, the object style compensation surpasses state-of-the-art methods, demonstrating its effectiveness.
Researcher Affiliation Academia Tingliang Feng1,2 Hao Shi1,3 Xueyang Liu1 Wei Feng1,2 Liang Wan1 Yanlin Zhou4 Di Lin1 1College of Intelligence and Computing, Tianjin University 2Laboratory of Computation and Analytics of Complex Management Systems (CACMS), Tianjin University 3Department of Automation, Tsinghua University 4Dunhuang Academy
Pseudocode No The paper describes the method using diagrams and text, but does not include any pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing code or a link to a code repository.
Open Datasets Yes We use GTA5 [57], SYNTHIA [58], CDriving [7], ACDC [12], Cityscapes [13], KITTI [14], and Wild Dash [15] datasets to evaluate our method.
Dataset Splits No All images with annotations in source domain and a portion of images without annotations in target domain are used for network training. We evaluate the segmentation performances on the rest images in target domain and all images in open domain.
Hardware Specification Yes The network is trained on a single RTX 3090 GPU.
Software Dependencies No We employ the Py Torch toolkit to implement our segmentation network with Object-Level Discrepancy Memory (OLDM).
Experiment Setup Yes To optimize the network, we utilize the SGD solver for 250,000 iterations. The initial learning rate is set to 0.00025, which undergoes a linear decay throughout the training process. Each mini-batch consists of 4 images, comprising 2 source images and 2 target images. We apply random cropping, flipping, color jittering, and Gaussian blurring techniques to the training images, using a crop size of 640 × 360.