Learning Deep Relations to Promote Saliency Detection

Authors: Changrui Chen, Xin Sun, Yang Hua, Junyu Dong, Hongwei Xv10510-10517

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments demonstrate that our method can significantly boost the performance of the state-of-the-art saliency detectors on various benchmark datasets. 5 Experiments
Researcher Affiliation Academia 1Ocean University of China, 2Queen s University Belfast
Pseudocode No The paper describes the method using prose and a diagram, but no structured pseudocode or algorithm block is provided.
Open Source Code No The code will be published on https://github.com/ouc-ocean-group/LDPS soon.
Open Datasets Yes In our experiment, we use four well-known saliency benchmark datasets to evaluate our method. HKU-IS (Li and Yu 2016) contains 4447 images with multiple salient objects. DUT-OMRON (Yang et al. 2013) includes 5168 complicated images with one or two salient objects. Pascal-S (Li et al. 2014) which contains 850 natural images is a subset of the PASCAL VOC2010 dataset. ECSSD (Yan et al. 2013) contains 1000 images with multiple objects of varying sizes. The training dataset of our model is DUTS-TE (Wang et al. 2017a) which has 5019 images collected from the Image Net DET dataset (Deng et al. 2009).
Dataset Splits No The paper mentions datasets used for training and testing (DUTS-TE for training, DUTS-TEST for training their model, and HKU-IS, DUT-OMRON, Pascal-S, ECSSD for evaluation), but does not provide specific percentages or counts for train/validation/test splits, nor does it explicitly mention a validation set.
Hardware Specification Yes The training is operated on a PC with a GTX 2080ti GPU. We use a single GTX 1080 during the inference phase.
Software Dependencies No The paper mentions Mobile Net v2 as a feature extractor but does not provide specific version numbers for software dependencies such as deep learning frameworks or other libraries.
Experiment Setup Yes All the thresholds for generating the masks are set as: θfg = 0.9, θa fg = 0.8, θa bg = 0.3. We train our model on only 1 GPU for 20k iterations, with a learning rate of 5e-4 for backbone and 5e-3 for the rest components. The learning rates are decreased by the polynomial learning rate policy.