Discriminatively Trained Multiscale Deformable Part Models

Version 1
This is an old release
The latest release is available here

Below is an implementation of the object detection system described in the paper:

A Discriminatively Trained, Multiscale, Deformable Part Model
Pedro Felzenszwalb, David McAllester and Deva Ramanan
IEEE Conference on Computer Vision and Pattern Recognition, 2008.
pdf

Currently we are distributing the object detection code with models trained on the PASCAL VOC 2006 dataset.
Models trained on the 2007 dataset will be available soon. The learning code will be also be available in the future.

The system is implemented in matlab, with a few helper functions implemented in C++ for efficiency reasons.
The README file has instructions on how to compile and use the system.

This software is released under the MIT License.
To download the source code, click here.


Recognition results in the PASCAL VOC 2006 benchmark

The models included with the source code were trained on the train+val dataset and evaluated on the test dataset.
This is exactly the protocol of the "comp3" competition.

Average precision (AP)
bicycle bus car cat cow dog horse motorbike person sheep
0.592 0.407 0.545 0.081 0.346 0.070 0.283 0.485 0.322 0.303
The PR curves are available here


Examples:

car model person model bicycle model