Home

Awesome

This is code to generate the bottom up segmentation, amodal completion, semantic segmentation and scene classification results for our CVPR 13 paper, Perceptual Organization and Recognition of Indoor Scenes from RGB-D Images. Saurabh Gupta, Pablo Arbelaez, Jitendra Malik, In CVPR 2013

Notes

  1. External Dependencies. Please put them in the folder external and modify startup.m and COM/getPaths.m

a. gPb-UCM, Contour and region benchmarking code from http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_full.tgz We modified the files match_segmentations2.m, ucm_mean_pb.cpp, buildW.cpp, and the patched files and patches are available in external/

b. VLFeat from http://www.vlfeat.org/

c. liblinear from http://www.csie.ntu.edu.tw/~cjlin/liblinear/, compile this to generate train.mex*, predict.mex*

d. liblinear dense version from http://ttic.uchicago.edu/~smaji/projects/digits/, compile this to generate trainDense.mex* and predictDense.mex*

e. SIFT color desprictors from http://www.colordescriptors.com, We used v2.1.

f. Image Stack Library https://code.google.com/p/imagestack/, for computing the local deph boundary cues.

  1. The code is intended for use with the data available here. This contains the color images, preprocessed point clouds and the preprocessed ground truth. https://www.dropbox.com/s/qjw05vyvhvzznee/data.tgz?dl=0. Please modify COM/getPaths.m and allBenchmarks/benchmarkPaths.m to point to where you put the data folder.

  2. We also provide the precomputed results, https://www.dropbox.com/s/kwge19n42k2r9ba/output.tgz?dl=0, and the pretrained models, https://www.dropbox.com/s/szcyjrfjhbz4vt8/model.tgz?dl=0.

Data Preprocessing

  1. The data that we provide is different from the one provided by Silberman etal in the following ways:

a. We use I(46:470, 41:600, :) part of the image. This is different from what they use by 1 pixel on each side, I think.

b. We use the original point clouds from the Kinect from over here http://horatio.cs.nyu.edu/mit/silberman/nyu_depth_v2/orig_images.mat, and project them to the RGB frame but instead of just keeping the Z coordinate also keep the X and the Y coordinate (keeping just the Z and projecting back rays to get X and Y introduces a little bit of error, but should still work well, we did not try using that...)

c. Finally, we preprocessed the ground truth to get rid of the double contours.

Precomputed Results

  1. https://www.dropbox.com/s/kwge19n42k2r9ba/output.tgz?dl=0: includes the folder cachedir/release/output which contains the UCM, amodal completion and the semantic segmentation outputs. Their are also .mat files with the performance of our system.

Pre trained Models

  1. https://www.dropbox.com/s/szcyjrfjhbz4vt8/model.tgz?dl=0, includes the folder cachedir/release/model which has the pretrained ucm models, various semantic segmentation and scene classification models.

Questions and Comments

Please feel free to email, sgupta at eecs dot berkeley dot edu for any questions and comments.