Weakly-Supervised Salient Object Detection Using Point Supervision
Authors: Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, Wenqiang Zhang670-678
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on five largest benchmark datasets demonstrate our method outperforms the previous state-of-the-art methods trained with the stronger supervision and even surpass several fully supervised state-of-the-art models. |
| Researcher Affiliation | Academia | 1Shanghai Key Laboratory of Intelligent Information Processing, School of Computer Science, Fudan University 2Academy for Engineering & Technology, Fudan University {sygao18,weizh,yanwang19,qyguo20,clzhang20,yjhe20,wqzhang}@fudan.edu.cn |
| Pseudocode | Yes | Algorithm 1: Flood Filling Algorithm Input: Seed point (x, y), image I, seted value α, old value I(x, y) Output: Filled mask M 1: flood filling algorithm ((x,y), I, α) 2: if x 0 and x < width and y 0 and y < height 3: and a < I(x, y) old < b and I(x, y) = α then 4: M(x,y) α 5: flood filling ((x + 1, y), I, α) 6: flood filling ((x 1, y), I, α) 7: flood filling ((x, y + 1), I, α) 8: flood filling ((x, y 1), I, α) 9: end if |
| Open Source Code | Yes | The code is available at: https://github.com/shuyonggao/PSOD. |
| Open Datasets | Yes | To minimize the labeling time consumption while providing location information of salient objects, we build a Point-supervised Dataset (P-DUTS) by relabeling DUTS (Wang et al. 2017) dataset, a widely used saliency detection dataset containing 10553 training images. ... To evaluate the performance, we experiment on five public used benchmark datasets: ECSSD (Yan et al. 2013), PASCAL-S (Li et al. 2014), DUT-O (Yang et al. 2013), HKU-IS (Li and Yu 2015), and DUTS-test. |
| Dataset Splits | No | The paper states that the P-DUTS dataset is used as the training set and evaluates on separate benchmark test datasets. However, it does not explicitly provide specific train/validation/test dataset splits (e.g., percentages or counts) or mention a dedicated validation set for hyperparameter tuning. |
| Hardware Specification | Yes | We train on four TITAN Xp GPUs. |
| Software Dependencies | No | The paper states the model is "implemented on the Pytorch toolbox" but does not provide a specific version number for Pytorch or any other software dependencies. |
| Experiment Setup | Yes | The maximum learning rate is set to 2.5 10 4 for the transformer part and 2.5 10 3 for other parts. Warm-up and linear decay strategies are used to adjust the learning rate. Stochastic gradient descent (SGD) is used to train the network, and the following hyper-parameters are used: momentum=0.9, weight decay=5 10 4. Horizontal flip and random crop are used as data augmentation. The batch size is set to 28 and it takes 20 epochs for the first training procedure. The hyperparameter γ of Eq. 1 is set to be 5. The second round of training uses the same parameters, but the masks are replaced with refined ones. |