Active Learning for Semantic Segmentation with Multi-class Label Query
Authors: Sehyun Hwang, Sohyun Lee, Hoyoung Kim, Minhyeon Oh, Jungseul Ok, Suha Kwak
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed framework achieved the state of the art on both Cityscapes [15] and PASCAL VOC 2012 [21]. Especially, it achieved 95% of the fully supervised learning performance on Cityscapes with only 4% of the full labeling cost. In addition, we verified the efficacy and efficiency of multi-class labeling through extensive empirical analyses: Its efficacy was demonstrated by experiments with varying datasets, model architectures, acquisition functions, and budgets, while its efficiency was examined in real-world annotation scenarios by measuring actual labeling time across a large number of human annotators. |
| Researcher Affiliation | Academia | Pohang University of Science and Technology (POSTECH), South Korea {sehyun03, lshig96, cskhy16, minhyeonoh, jungseul.ok, suha.kwak}@postech.ac.kr |
| Pseudocode | No | The paper describes the steps of its algorithms but does not provide structured pseudocode or explicitly labeled algorithm blocks. |
| Open Source Code | Yes | Our code and results are available at https://github.com/sehyun03/MulActSeg. |
| Open Datasets | Yes | Our method is evaluated on two semantic segmentation datasets, Cityscapes [15] and PASCAL VOC 2012 (VOC) [21]. |
| Dataset Splits | Yes | The former contains 2975 training, 500 validation, and 1525 test images with 19 semantic classes. The latter consists of 1464 training and 1449 validation images with 20 semantic classes. We evaluated models on validation splits of these datasets. |
| Hardware Specification | No | The paper mentions 'We adopt Deep Labv3+ [13] with Res Net-50/101 pretrained on Image Net [18] as our segmentation models', but does not provide specific hardware details like GPU/CPU models, processors, or memory used for experiments. |
| Software Dependencies | No | The paper states 'We implement our method using the Py Torch framework [51]' and mentions 'Adam W [45] for optimization', but it does not provide specific version numbers for these software components or any other libraries. |
| Experiment Setup | Yes | The balancing hyperparameters λCE and λMP of Eq. (10) are set to 16 and 8, respectively, and the temperature τ was fixed by 0.1. In both datasets we utilize 32 × 32 superpixel regions given by SEEDS [59]. For Cityscapes, initial learning rates are set to 2e-3 (stage 1) and 4e-3 (stage 2), and ν in Eq. (4) is set to 6. The models are trained for 80K iterations with mini-batches of four 769 × 769 images. For VOC, we configure ν to 12 and train the models for 30K iterations using a learning rate of 1e-3 in both stages. Each mini-batch consists of twelve 513 × 513 images. More details are given in the Appendix B. |