Pushing the Boundaries of Boundary Detection using Deep Learning

Authors: Iasonas Kokkinos

ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental When measured on the standard Berkeley Segmentation Dataset, we improve theoptimal dataset scale F-measure from 0.780 to 0.808 while human performance is at 0.803.
Researcher Affiliation Academia Iasonas Kokkinos Center for Visual Computing Centrale Sup elec and INRIA Chatenay-Malabry, 92095, France {iasonas.kokkinos}@ecp.fr
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper states 'Our detector is fully integrated in the popular Caffe framework' but does not explicitly state that the code for their method is open-source or provide a link.
Open Datasets Yes When measured on the standard Berkeley Segmentation Dataset, we improve theoptimal dataset scale F-measure from 0.780 to 0.808 while human performance is at 0.803. We have not used these additional scalings in our experiments due to time constraints, but have considered the use of boundaries from the VOC Context dataset (Mottaghi et al., 2014)
Dataset Splits No The paper mentions the 'BSD trainval set' and 'BSD test set' but does not provide specific details on the percentages or counts for training, validation, and test splits.
Hardware Specification Yes All stages are implemented in Caffe, requiring less than a second on an Nvidia Titan GPU.
Software Dependencies No The paper mentions 'Caffe framework' and 'Damascene system' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Our contributions consist firstly in combining a careful design of the loss for boundary detection training, a multi-resolution architecture and training with external data... we modify the training objective by associating the side term with a temporally decreasing weight while keeping the second term s weight fixed: L(t)(W, w, h) = (1 t/T)Lside(W, w) + Lfuse(W, w, h)... we consider a DSN-type multiresolution architecture with tied weights, meaning that layers that operate at different resolutions share weights with each other.