Few-Shot Segmentation via Cycle-Consistent Transformer

Authors: Gengwei Zhang, Guoliang Kang, Yi Yang, Yunchao Wei

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on all few-shot segmentation benchmarks demonstrate that our proposed Cy CTR leads to remarkable improvement compared to previous state-of-the-art methods. Specifically, on Pascal-5i and COCO-20i datasets, we achieve 67.5% and 45.6% m Io U for 5-shot segmentation, outperforming previous state-of-the-art method by 5.6% and 7.1% respectively.
Researcher Affiliation Collaboration Gengwei Zhang1,2 , Guoliang Kang3, Yi Yang4, Yunchao Wei5,6 1 Baidu Research 2 Re LER, Centre for Artificial Intelligence, University of Technology Sydney 3 University of Texas, Austin 4 CCAI, College of Computer Science and Technology, Zhejiang University 5 Institute of Information Science, Beijing Jiaotong University 6 Beijing Key Laboratory of Advanced Information Science and Network
Pseudocode No The paper does not contain any explicit 'Pseudocode' or 'Algorithm' sections, nor does it present structured code-like blocks.
Open Source Code No The paper does not provide any links to open-source code for the described methodology or state that the code is publicly available.
Open Datasets Yes We conduct experiments on two commonly used few-shot segmentation datasets, Pascal-5i [10] (which is combined with SBD [11] dataset) and COCO-20i [17], to evaluate our method.
Dataset Splits Yes For Pascal-5i, 20 classes are separated into 4 splits. For each split, 15 classes are used for training and 5 classes for test. At the test time, 1,000 pairs that belong to the testing classes are sampled from the validation set for evaluation. In COCO-20i, we follow the data split settings in FWB [23] to divide 80 classes evenly into 4 splits, 60 classes for training and test on 20 classes, and 5,000 validation pairs from the 20 classes are sampled for evaluation.
Hardware Specification Yes Experiments are carried out on Tesla V100 GPUs.
Software Dependencies No The paper mentions optimizers (SGD, Adam W) and loss functions (Dice loss) but does not provide specific version numbers for any software libraries or dependencies used in the implementation.
Experiment Setup Yes In our experiments, the training strategies follow the same setting in [30]: training for 50 epochs on COCO-20i and 200 epochs on Pascal-5i. Images are resized and cropped to 473 473 for both datasets and we use random rotation from 10 to 10 as data augmentation. Besides, we use Image Net [25] pretrained Res Net [12] as the backbone network and its parameters (including Batch Norms) are freezed. For the parameters except those in the transformer layers, we use the initial learning rate 2.5 10 3, momentum 0.9, weight decay 1 10 4 and SGD optimizer with poly learning rate decay [4]. The mini batch size on each gpu is set to 4.