[Re] Interactive Two-Stream Decoder for Accurate and Fast Saliency DetectionDownload PDF

06 Dec 2020 (modified: 05 May 2023)ML Reproducibility Challenge 2020 Blind SubmissionReaders: Everyone
Abstract: This article describes our attempts to reproduce the work presented in the paper - Interactive Two-Stream Decoder for Accurate and Fast Saliency Detection which published in CVPR2020, as part of the ML Reproducibility Challenge 2020. The work exploited the correlation between saliency and contour information by introducing interactive connections between the saliency and contour streams in the decoder. The interactive connections allow the saliency and contour information from the respective stream to be transmitted across the branch and therefore obliging the network to learn the correlation between saliency and contour information. Our work consists of three parts: (1) evaluation using the provided pre-trained models from the original paper; (2) evaluation using our retrained models; (3) parameters analysis. Despite differences in experimental results, the reproducibility of the studied paper is further ascertained by conducting the statistical test.
Paper Url: https://openreview.net/forum?id=TzLBdfuy814&noteId=QpFxzuZX7Wm
4 Replies

Loading