Learning Saliency-Free Model with Generic Features for Weakly-Supervised Semantic Segmentation
Authors: Wenfeng Luo, Meng Yang11717-11724
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on the PASCAL VOC 2012 dataset show that the proposed saliency-free method outperforms the previous approaches under the same weakly-supervised setting and achieves superior segmentation results, which are 64.5% on the validation set and 64.6% on the test set concerning m Io U metric. |
| Researcher Affiliation | Academia | Wenfeng Luo,1 Meng Yang1,2 1School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China 2Key Laboratory of Machine Intelligence and Advanced Computing (SYSU), Ministry of Education |
| Pseudocode | Yes | Algorithm 1: Estimating proxy ground-truth |
| Open Source Code | No | The result on test set is available on the website (http://host.robots.ox.ac.uk:8080/ anonymous/USWTK1.html). This link is for results, not code, and no other statements about code availability are present. |
| Open Datasets | Yes | PASCAL VOC 2012 (Everingham et al. 2012) and COCO (Lin et al. 2014) |
| Dataset Splits | Yes | We report the mean intersection over union (m Io U) on both val and test set. COCO: We use the train-val split setting of competition in 2017, where 112k images are used for training and the remaining 5k are reserved for evaluation. |
| Hardware Specification | No | The paper does not explicitly state the specific hardware used to run its experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions software components and models (e.g., 'Adam optimizer', 'Deep Lab-CRF-Large FOV model', 'Res Net-101') but does not provide specific version numbers for any libraries or frameworks used. |
| Experiment Setup | Yes | We use Adam optimizer (Kingma and Ba 2014) with a learning rate of 5e-6 for the backbone and 1e-4 for the randomly initialized layers. With a batch size of 16, we train the segmentation network for 20 epochs. During training, the image batches are resized to fixed dimension of 328 328. In the test phase, we adopt multi-scale testing with input dimensions of 241, 328 and 401. |