Welcome to the Exemplar-SVM library, a large-scale object recognition library developed at Carnegie Mellon University while obtaining my PhD in Robotics. -- Tomasz Malisiewicz
The code is written in Matlab and is the basis of the following two projects, as well as my doctoral dissertation:
- Tomasz Malisiewicz, Abhinav Gupta, Alexei A. Efros. Ensemble of Exemplar-SVMs for Object Detection and Beyond. In ICCV, 2011. PDF | Project Page
- Abhinav Shrivastava, Tomasz Malisiewicz, Abhinav Gupta, Alexei A. Efros. Data-driven Visual Similarity for Cross-domain Image Matching. In SIGGRAPH ASIA, December 2011. PDF | Project Page
More details and experimental evaluation can be found in my PhD thesis, available to download as a PDF.
This object recognition library uses some great open-source software:
Linear SVM training: libsvm-3.0-1
Fast blas convolution code (from voc-release-4.0),
HOG feature code (31-D) (from voc-release-3.1),
VOC development/evaluation code imported from the PASCAL VOC website
MATLAB Quick Start Guide
To get started, you need to install MATLAB and download the code from Github. This code has been tested on Mac OS X and Linux. Pre-compiled Mex files for Mac OS X and Linux are included.
Download Exemplar-SVM Library source code (MATLAB and C++) and compile it
$ cd ~/projects/ $ git clone email@example.com:abhi2610/exemplarsvm.git $ cd ~/projects/exemplarsvm $ matlab >> esvm_compile
Quick Guide for SIGGRAPH Asia paper
Demo Script: Training exemplar-SVM for one image
See the demo walk-through tutorial/esvm_demo_single_exemplar_training.html for instructions on training Exemplar-SVMs for single images.
Function Description: Training exemplar-SVM for one image
See the detailed explanation tutorial/esvm_train_single_exemplar.html of the function for training Exemplar-SVMs for single images.
Demo: Example of applying models to a single image or a set of images
See the demo walk-through tutorial/esvm_demo_apply.html for a step-by-step tutorial on applying Exemplar-SVMs to images. Or you can just run the demo:
Quick Guide for ICCV paper for PASCAL Object Detection Task
Download and load pre-trained VOC2007 model(s)
$ matlab >> addpath(genpath(pwd)) >> [models, M, test_set] = esvm_download_models('voc2007-bus');
$ wget http://people.csail.mit.edu/~tomasz/exemplarsvm/voc2007-models.tar $ tar -xf voc2007-models.tar $ matlab >> load voc2007_bus.mat >> [models, M, test_set] = esvm_download_models('voc2007-bus.mat');
You can alternatively download the pre-trained models individually from http://people.csail.mit.edu/tomasz/exemplarsvm/models/ or a tar file of all models voc2007-models.tar (NOTE: tar file is 450MB)
Training an Ensemble of Exemplar-SVMs
Toy Demo: Exemplar-SVM training and testing on a set of synthetic images
See the synthetic training demo walk-through tutorial/esvm_demo_train_synthetic.html for a step-by-step tutorial on how to set-up images and bounding boxes for a training experiment. Or you can run the synthetic training demo:
The training scripts are designed to work with the PASCAL VOC 2007 dataset, so we need to download that first.
Prerequsite: Install PASCAL VOC 2007 trainval/test sets
$ mkdir /nfs/baikal/tmalisie/pascal #Make a directory for the PASCAL VOC data $ cd /nfs/baikal/tmalisie/pascal $ wget http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2007/VOCtrainval_06-Nov-2007.tar $ wget http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2007/VOCtest_06-Nov-2007.tar $ tar xf VOCtest_06-Nov-2007.tar $ tar xf VOCtrainval_06-Nov-2007.tar
Demo: Training and Evaluating an Ensemble of "bus" Exemplar-SVMs quick-demo
>> data_dir = '/your/directory/to/pascal/VOCdevkit/'; >> dataset = 'VOC2007'; >> results_dir = '/your/results/directory/'; >> [models,M] = esvm_demo_train_voc_class_fast('car', data_dir, dataset, results_dir); # All output (models, M-matrix, AP curve) has been written to results_dir
See the file tutorial/esvm_demo_train_voc_class_fast.html for a step-by-step tutorial on what esvm_demo_train_voc_class_fast.m produces
Script: Training and Evaluating an Ensemble of "bus" Exemplar-SVMs full script
>> data_dir = '/your/directory/to/pascal/VOCdevkit/'; >> dataset = 'VOC2007'; >> results_dir = '/your/results/directory/'; >> [models,M] = esvm_script_train_voc_class('bus', data_dir, dataset, results_dir); # All output (models, M-matrix, AP curve) has been written to results_dir
Extra: How to run the Exemplar-SVM framework on a cluster
This library was meant to run on a cluster with a shared NFS/AFS file structure where all nodes can read/write data from a common data source/target. The PASCAL VOC dataset must be installed on such a shared resource and the results directory as well. The idea is that results are written as .mat files and intermediate work is protected via lock files. Lock files are temporary files (they are directories actually) which are deleted once something has finished process. This means that the entire voc training script can be replicated across a cluster, you can run the script 200x times and the training will happen in parallel.
To run ExemplarSVM on a cluster, first make sure you have a cluster, use an ssh-based launcher such as my warp_scripts github project. I have used warp_starter.sh at CMU (using WARP cluster) and sc.sh at MIT (using the continents).
Here is the command I often use at MIT to start Exemplar-SVM runs, where machine_list.sh contains computer names
$ cd ~ $ git clone firstname.lastname@example.org:quantombone/warp_scripts.git $ cd ~/warp_scripts/ $ cp machine_list.sh-example machine_list.sh $ nano machine_list.sh #now edit the file to point to your cluster CPUs $ ./sc.sh "cd ~/projects/exemplarsvm; addpath(genpath(pwd)); esvm_script_train_voc_class('train');"
Updated by Abhinav Shrivastava
Copyright (C) 2011 by Tomasz Malisiewicz
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.