This software package contains code for the discriminative framework described in our ECCV 2012 paper, "Undoing the Damage of Dataset Bias". The provided code is for the classification experiments described in the paper, and has only been tested on Linux.
Before you can use the code, you need to download this repository and compile the learning code:
$ git clone http://github.com/adikhosla/undoing-bias
$ cd undoing-bias
$ make
There are two demo scripts provided in this code to run the classification experiments from the paper: demo_small.m and demo_full.m. demo_small executes the learning code on randomly generated data and completes execution in under 5 seconds. demo_full executes the learning code on the full set of features (details below), and completes execution in about 8 minutes.
>> demo_small
>> demo_full
The demo above will output AP values on the unseen test set, as well as AP values on the test partition of the seen sets.
We used 4 datasets in our experiments, namely Caltech-101, LabelMe, PASCAL VOC 2007 and SUN09 (described in the paper). The features used in the classification experiments are available for download from the project website in Matlab format.
The archive contains 4 mat files with features for the train and test images and labels for the five object categories used in our experiments (car, cat, chair, dog, person).
We have modified the code from the discriminatively trained deformable parts model (voc-release3) to implement our max-margin framework, and included some auxillary files with no modifications. In addition, we adapt code from the PASCAL VOC development kit for Average Precision (AP) calculation.
If you have any feedback, please email Aditya Khosla at [email protected].
Please cite our paper if you use this code:
Undoing the Damage of Dataset Bias
Aditya Khosla, Tinghui Zhou, Tomasz Malisiewicz, Alexei A. Efros, Antonio Torralba.
European Conference on Computer Vision (ECCV), 2012