Unsupervised Representation for Semantic Segmentation by Implicit Cycle-Attention Contrastive Learning
Authors: Bo Pang, Yizhuo Li, Yifan Zhang, Gao Peng, Jiajun Tang, Kaiwen Zha, Jiefeng Li, Cewu Lu2044-2052
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate it on PASCAL-VOC (Everingham et al. 2010) and Cityscapes (Cordts et al. 2016) datasets under linear protocol setting to verify results linear separability. Extensive experiments show that CACL works robustly and we observe consistent performance improvements over baselines. |
| Researcher Affiliation | Academia | 1 Shanghai Jiao Tong University 2 Massachusetts Institute of Technology |
| Pseudocode | No | The paper describes methods in paragraph text and equations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any statement or link indicating that the source code for the methodology is openly available. |
| Open Datasets | Yes | The unsupervised training is conducted on Image Net (Deng et al. 2009) and Kinetics-400 (Carreira and Zisserman 2017) datasets. The experimental evaluations are conducted on two commonly used benchmarks, PASCAL-VOC (Everingham et al. 2010) and Cityscapes (Cordts et al. 2016) datasets. |
| Dataset Splits | No | The paper mentions 'The training and evaluation splits in PASCAL-VOC contain about 10k augmented images covering 20 classes, and 3.5k images covering 19 classes in Cityscapes', but does not provide specific train/validation/test percentages or explicit counts for each split, nor does it explicitly mention a validation set. |
| Hardware Specification | No | The paper mentions '16 GPU' but does not specify the particular models or types of GPUs, CPUs, or other hardware used for the experiments. |
| Software Dependencies | No | The paper mentions software components like 'LARS', 'SGD', 'Mo Co-v2', and 'Res Net' but does not provide specific version numbers for these or other software dependencies such as deep learning frameworks or CUDA versions. |
| Experiment Setup | Yes | The optimizer we adopt for training is LARS (You, Gitman, and Ginsburg 2017) with SGD. The momentum is set to 0.9, while 1e 6 for weight decay and 0.001 for the trust coefficient of LARS. The initial learning rate is 1.0 and decayed with a cosine schedule scheme (Loshchilov and Hutter 2016). The contrastive dimension and temperature τ are 128 and 0.07 just as Mo Co (He et al. 2020). |