Salient Object Detection via Augmented Hypotheses

Authors: Tam Van Nguyen, Jose Sepulveda

IJCAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We finally evaluate the proposed framework on two challenging datasets, MSRA1000 and i Co Seg. Our extensive experimental results show that our method outperforms state-of-the-art approaches.
Researcher Affiliation Academia Tam V. Nguyen, Jose Sepulveda Department for Technology, Innovation and Enterprise Singapore Polytechnic {nguyen van tam, sepulveda jose}@sp.edu.sg
Pseudocode Yes Algorithm 1 Superpixel compactness computation
Open Source Code No The paper does not provide any links to open-source code or explicitly state that the code for the methodology is available.
Open Datasets Yes We evaluate and compare the performances of our algorithm against previous baseline algorithms on two representative benchmark datasets: the MSRA 1000 salient object dataset [Achanta et al., 2009] and the Interactive cosegmentation Dataset (i Co Seg) [Batra et al., 2010].
Dataset Splits No The paper mentions evaluating precision/recall curves and adaptive thresholds, but does not explicitly describe train/validation/test dataset splits for their own model training. It mentions that the BING hypothesis generator was trained on VOC2007, but this is a component they use, not their own model's data split.
Hardware Specification Yes The average time of each method is measured on a PC with Intel i7 3.3 GHz CPU and 8GB RAM.
Software Dependencies No The paper mentions using BING, SLIC, C++, and MATLAB, but does not provide specific version numbers for these software components.
Experiment Setup Yes BING hypothesis generator is trained with VOC2007 dataset [Everingham et al., 2010] same as in [Cheng et al., 2014]. In order to compute the foreground map, θ is set as 0.1 and we convert the color channels from RGB to Lab color space as suggested in [Achanta et al., 2009; Perazzi et al., 2012]. Regarding the image over-segmentation, we use SLIC [Achanta et al., 2012] for the superpixel segmentation. We set the number of superpixels as 100 as a trade-off between the fine oversegmentation and the processing time.