Consistent-Separable Feature Representation for Semantic Segmentation

Authors: Xingjian He, Jing Liu, Jun Fu, Xinxin Zhu, Jinqiao Wang, Hanqing Lu1531-1539

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

Reproducibility Variable Result LLM Response
Research Type Experimental To verify the effectiveness of the proposed method, we plug our module into recent state-of-the-arts segmentation methods (e.g., Non-local, PSP, Deeplabv3+.), and the results indicate that our method achieves impressive improvements compared with these strong baselines. Our proposed approach achieves state-of-the-art performance on Cityscapes (82.6% m Io U in test set), ADE20K (46.65% m Io U in validation set), COCO Stuff (41.3% m Io U in validation set) and PASCAL Context (55.9% m Io U in test set).
Researcher Affiliation Academia 1School of Artificial Intelligence, University of Chinese Academy of Sciences 2National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions that 'The Py Torch framework is employed to implement our network' but does not provide any link or explicit statement about the availability of their source code.
Open Datasets Yes We further carry out extensive experiments on four competitive datasets, including Cityscapes dataset (Cordts et al. 2016), ADE20K dataset (Zhou et al. 2017), COCO Stuff dataset (Caesar, Uijlings, and Ferrari 2018), and PASCAL Context dataset (Mottaghi et al. 2014).
Dataset Splits Yes Cityscapes It contains 5000 high quality pixel-level annotated images. And these annotated images are divided into 2975, 500 and 1525 images for training, validation, and testing. ADE20K There are 20K images in the training set, 2K images in the validation set, and 3K images in the testing set.
Hardware Specification No The paper mentions using PyTorch but does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'The Py Torch framework is employed to implement our network' but does not specify its version number or any other software dependencies with version details.
Experiment Setup Yes During training phase, the initial learning rate is set to 0.01 for the Cityscapes dataset with the momentum of 0.9, the weight decay of 0.0001 and the batch size of 8. Following (Zhang et al. 2018), the poly learning policy is used to decay the initial learning rate by multiplying (1 iter total iter)0.9 after each iteration. For data augmentation, random horizontal filp, random cropping (cropsize 769) and random resizing with scales range [0.75, 2.0] are employed in the ablation study. Besides, we train the model with Synchronized Batch Normalization for 40k iterations for ablation study, and 90k iterations for submission to the server. During the testing phase, following (Zhang et al. 2018), the sliding-window method is used for evaluation. ...we set λ1 = 1, λ2 = λ3 = 0.2.