Promoting Saliency From Depth: Deep Unsupervised RGB-D Saliency Detection

Authors: Wei Ji, Jingjing Li, Qi Bi, chuan guo, Jie Liu, Li Cheng

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superior efficiency and effectiveness of our approach in tackling the challenging unsupervised RGB-D SOD scenarios. Moreover, our approach can also be adapted to work in fully-supervised situation. Empirical studies show the incorporation of our approach gives rise to notably performance improvement in existing supervised RGB-D SOD models.
Researcher Affiliation Academia Wei Ji1, Jingjing Li1, , Qi Bi2, Chuan Guo1, Jie Liu3, Li Cheng1 1University of Alberta, Canada 2Wuhan University, China 3Dalian University of Technology, China
Pseudocode No The paper describes the system architecture and its components through text and diagrams (e.g., Figure 2, Figure 7), but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Implementation details are presented in the Appx A.3. Source code is publicly available.
Open Datasets Yes NJUD (Ju et al., 2014) in its latest version consists of 1,985 samples, that are collected from the Internet and 3D movies; NLPR (Peng et al., 2014) has 1,000 stereo images collected with Microsoft Kinect; STERE (Niu et al., 2012) contains 1,000 pairs of binocular images downloaded from the Internet; DUTLF-Depth (Piao et al., 2019) has 1,200 real scene images captured by a Lytro2 camera.
Dataset Splits No The paper specifies the construction of training and testing sets, but it does not explicitly mention or detail a separate validation set split (e.g., for hyperparameter tuning) or a cross-validation setup. It states: 'The remaining images are reserved for testing.'
Hardware Specification Yes The proposed deep unsupervised pipeline is implemented with Py Torch and trained using a single Tesla P40 GPU.
Software Dependencies No The paper states that the pipeline is 'implemented with Py Torch,' but it does not provide specific version numbers for PyTorch or any other software libraries or dependencies required for reproducibility.
Experiment Setup Yes All training & testing images are uniformly resized to the size of 352 352. Throughout training, the learning rate is set to 1 10 4, and the Adam optimizer is used with a mini-batch size of 10. Our approach is trained in an unsupervised manner, i.e., without any human annotations, where initial pseudo-labels are the outputs of handcrafted RGB-D saliency method CDCP (Zhu et al., 2017b). As for the proposed attentive training strategy, its alternation interval is set to 3, amounting to 2τ = 6 epochs in a training round.