Deep Edge-Aware Filters

Authors: Li Xu, Jimmy Ren, Qiong Yan, Renjie Liao, Jiaya Jia

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

Reproducibility Variable Result LLM Response
Research Type Experimental We use our method to simulate a number of practical operators, including but not limited to bilateral filter (BLF) (Paris & Durand, 2006), local Laplacian filter (LLF) (Paris et al., 2011), region covariance filter (Reg Cov) (Karacan et al., 2013), shock filter (Osher & Rudin, 1990), weighted least square (WLS) smoothing (Farbman et al., 2008), L0 smoothing (Xu et al., 2011), weighted median filter (Zhang et al., 2014b), rolling guidance filter (Zhang et al., 2014a), and RTV texture smoothing (Xu et al., 2012). They are representative as both localand globalschemes are included and effect of smoothing, sharpening, and texture removal can be produced. Our implementation is based on the Matlab VCNN framework (Ren & Xu, 2015).1We quantitatively evaluate our method (Paris et al., 2011). The average PSNRs and SSIMs are plotted in Fig. 11. All are high to produce usable results.
Researcher Affiliation Collaboration Li Xu XULI@SENSETIME.COM Jimmy SJ. Ren RENSIJIE@SENSETIME.COM Qiong Yan YANQIONG@SENSETIME.COM Sense Time Group Limited Renjie Liao RJLIAO@CSE.CUHK.EDU.HK Jiaya Jia LEOJIA@CSE.CUHK.EDU.HK The Chinese University of Hong Kong
Pseudocode Yes Algorithm 1 Deep Edge-Aware Filters
Open Source Code Yes Our implementation is available at the project webpage http://lxu.me/projects/deaf.
Open Datasets No We randomly collect one million 64 64 patches from high resolution natural image data obtained from flickr and their smoothed versions as training samples.
Dataset Splits No No specific validation split percentages, counts, or methodology provided. The paper mentions 'one million 64x64 patches from high resolution natural image data obtained from flickr' for training and '100 testing images'.
Hardware Specification Yes Table 1. Running time for different resolution images on desktop PC (Intel i7 3.6GHz with 16GB RAM, Geforce GTX 780 Ti with 3GB memory).
Software Dependencies No Our implementation is based on the Matlab VCNN framework (Ren & Xu, 2015).
Experiment Setup Yes input: one million image patches {Ii}, learning rate η, regularization parameter λ; initialization: {Wn} N(0, 1), {bn} 0; learning rate decay η 0.001/(1 + i 1E 7) ;In our implementation, the convolution kernel is of the size 16 16 and k = 256.