CPRAL: Collaborative Panoptic-Regional Active Learning for Semantic Segmentation
Authors: Yu Qiao, Jincheng Zhu, Chengjiang Long, Zeyao Zhang, Yuxin Wang, Zhenjun Du, Xin Yang2108-2116
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on Cityscapes and BDD10K datasets and show that CPRAL outperforms the cutting-edge methods with impressive results and less labeling proportion. |
| Researcher Affiliation | Collaboration | 1Dalian University of Technology 2JD Finance America Corporation 3SIASUN Robot & Automation CO.,Ltd |
| Pseudocode | No | The paper describes its methodology and processes using text and diagrams (Figure 1 and Figure 2), but it does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions designing an interactive GUI tool but does not explicitly state that the source code for the described methodology or this tool is open-source or provide any repository link or explicit code release statement. |
| Open Datasets | Yes | Here we demonstrate the performance on two public semantic segmentation datasets: Cityscapes (Cordts et al. 2016) and BDD100K (Yu et al. 2020). |
| Dataset Splits | Yes | There are 5000 examples in Cityscapes, 2975 for training, 500 for validation, and 1525 for testing. We use the training set to train different models and verify their performance on the validation set. BDD100K consists of 7000 training images, 1000 validation images, and 2000 testing images. We also train the models on the training set and verify them on the validation set. |
| Hardware Specification | Yes | We implement CPRAL using Py Torch and Tesla P100 graphics cards. |
| Software Dependencies | No | The paper mentions the use of 'Py Torch' but does not provide specific version numbers for PyTorch or any other software dependencies required to replicate the experiment. |
| Experiment Setup | Yes | Random horizontal flip and Gaussian blur are employed to augment the sampling diversity. The segmentation and loss prediction modules adopt the (SGD) optimizer with a momentum of 0.9 and a weight decay of 0.0005. For each sampling iteration, there are 50 epochs for training. The initial learning rate is 0.001 and drops to 0.0001 at epoch 35. The sampling will iterate five times on Cityscapes with initial 200 examples and six times on BDD100K with initial 400 examples. The panoptic subset size is 400 for Cityscapes and 800 for BDD100K, and the final selection is 200 and 400. The batch size for Mobile Net and DRN are 4 and 2, separately. The loss function for the segmentation model is cross-entropy, and the panoptic loss is the same as (Yoo and Kweon 2019). |