Recursive Context Propagation Network for Semantic Scene Labeling

Authors: Abhishek Sharma, Oncel Tuzel, Ming-Yu Liu

NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on Stanford background and SIFT Flow datasets show that the proposed method outperforms previous approaches.
Researcher Affiliation Collaboration Abhishek Sharma University of Maryland College Park, MD bhokaal@cs.umd.edu Oncel Tuzel Ming-Yu Liu Mitsubishi Electric Research Labs (MERL) Cambridge, MA {oncel,mliu}@merl.com
Pseudocode No The paper describes its methods in prose and diagrams but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions using 'publicly available implementation CAFFE [10]' and a 'publicly available implementation1' for L-BFGS, but does not state that the code for their proposed method (rCPN) is open-source or available.
Open Datasets Yes We extensively tested the proposed model on two widely used datasets for semantic scene labeling Stanford background [13] and SIFT Flow [14].
Dataset Splits Yes We used the 572 train and 143 test image split provided by [7] for reporting the results. SIFT Flow contains 2688, 256 256 color images with 33 semantic classes. We experimented with the train/test (2488/200) split provided by the authors of [15].
Hardware Specification Yes Our fast method (Section 2.1) takes only 0.37 seconds (0.3 for super-pixel segmentation, 0.06 for feature extraction and 0.01 for r CPN and labeling) to label a 256 256 image starting from the raw RGB image on a GTX Titan GPU and 1.1 seconds on a Intel core i7 CPU.
Software Dependencies No The paper mentions using 'CAFFE [10]' and 'Limited memory BFGS [12]' with a link to an implementation, but does not specify version numbers for these or any other software dependencies used in their experiments.
Experiment Setup Yes All the training images were flipped horizontally to get twice the original images. We used dropout in the last layer with dropout ratio equal to 0.5. Standard back-propagation for CNN is used with stochastic gradient descent update scheme on mini-batches of 6 images, with weight decay (λ = 5 10 5) and momentum (µ = 0.9). ...dsem = 60 for all the experiments.