Multi-Type Self-Attention Guided Degraded Saliency Detection

Authors: Ziqi Zhou, Zheng Wang, Huchuan Lu, Song Wang, Meijun Sun13082-13089

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on seven widely-used datasets show that our approach produces good performance on both clear and degraded images. We compare our model with other 13 state-of-the-art CNN-based models, including AFNet (Feng, Lu, and Ding 2019), BASNet (Qin et al. 2019), CPDNet (Wu, Su, and Huang 2019), MWSNet (Zeng et al. 2019), C2SNet (Li et al. 2018), Pi CANet (Liu, Han, and Yang 2018), DGRL (Wang et al. 2018b), LFRNet (Zhang et al. 2018), RFCN (Wang et al. 2018a), Amulet (Zhang et al. 2017a), UCF (Zhang et al. 2017b), WSS (Wang et al. 2017), and MSRNet (Li et al. 2017).
Researcher Affiliation Academia 1College of Intelligence and Computing, Tianjin University, Tianjin, China 2School of Information and Communication Engineering, Dalian University of Technology, Dalian, China 3Department of Computer Science and Engineering, University of South Carolina, USA 4Peng Cheng Laboratory
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating the availability of open-source code for the described methodology.
Open Datasets Yes We employ DUTS-TR (Wang et al. 2017) to train, which contains 10,553 images, and we perform image enhancement, i.e., reverse, mirroring to increase to 65k. Seven large benchmarks are used for evaluation, including DUT-OMRON (Yang et al. 2013), ECSSD (Shi et al. 2016), HKU-IS (Li and Yu 2016), PASCAL-S (Li et al. 2014), SOD (Movahedi and Elder 2010), MSRA-B (Jiang et al. 2013), and DUTS-TE (Wang et al. 2017).
Dataset Splits No The paper mentions using DUTS-TR for training and DUTS-TE for testing, but does not specify validation splits or percentages for any dataset used.
Hardware Specification Yes All the experiments are run on NVIDIA Geforce GTX 1080Ti (11 GB memory) and i7-8700k cpu.
Software Dependencies No The paper states 'The whole architecture is built on the keras deep learning framework' but does not provide specific version numbers for Keras or any other software dependencies.
Experiment Setup Yes All training and testing images are resized to 224 224. We use the Adam optimizer to train and the learning rate is initialized to 1e-5. Ltotal = λ1Lbce + λ2Lssim + λ3LAT ... λ1, λ2 and λ3 is set to 5, 10 and 2 respectively. The first five convblocks of MSANet are all initialized from VGGNet16, and the other parameters are all randomly assigned.