Semantic-Aware Superpixel for Weakly Supervised Semantic Segmentation
Authors: Sangtae Kim, Daeyoung Park, Byonghyo Shim
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments 4.1 Datasets and Experiment Settings We evaluate the proposed approach on the PASCAL VOC 2012 (Everingham et al. 2015) and MS-COCO 2014 (Lin et al. 2014) segmentation benchmark datasets. 4.3 Ablation Studies In the proposed method, we use the feature obtained from the DINO. To investigate the effects of different features, we compare the superpixels generated using various features and discuss the qualities of the obtained superpixels. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, Seoul National University, Seoul, Korea 2 Department of Information and Communication, Inha University, Incheon, Korea |
| Pseudocode | No | The paper describes procedures in text and flowcharts (Figure 2) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/st17kim/semanticaware-superpixel. |
| Open Datasets | Yes | We evaluate the proposed approach on the PASCAL VOC 2012 (Everingham et al. 2015) and MS-COCO 2014 (Lin et al. 2014) segmentation benchmark datasets. |
| Dataset Splits | Yes | PASCAL VOC has 20 foreground classes and one background class and consists of 1,464 training images, 1,449 validation images, and 1,456 test images. As in many practices (Chen et al. 2017; Wei et al. 2017), additional dataset is augmented to training dataset, resulting 10,582 training images in total (Hariharan et al. 2011). MS-COCO has 80 foreground classes and one background class and consists of 82,783 training images and 40,504 validation images. |
| Hardware Specification | Yes | The classification network and the segmentation network are trained on a single NVIDIA Ge Force Titan Xp. |
| Software Dependencies | No | The paper states: 'Our approach is implemented with Tensorflow (Abadi et al. 2016).' It mentions Tensorflow but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | We set τ to 0.3 in superpixel discovery method. We set α = 0.6 to identify the foreground pixels and β = 0.7 for the criterion in seeded region growing. We use the stochastic gradient descent optimizer with the momentum 0.9. We set the weight decay to 0.0005 and the batch size to 20. We employ polynomial learning rate policy (Liu, Rabinovich, and Berg 2015) with initial learning rate 10 3 and power 0.9, i.e., L = 10 3 (1 iter/maxiter)0.9. In early training iterations, we gradually increase the learning rate from 10 6 to 10 3 through the first three epochs. The learning rate for the last layers is multiplied by 10. We train the segmentation network for 15 epochs. |