This repository provides an implementation of Deformable Part Models (DPMs) for object detection, specifically Release 5 of the system described in Ross Girshick's Ph.D. dissertation. It targets researchers and practitioners in computer vision needing a robust, discriminatively trained object detection system with support for latent SVM and weak-label structural SVM, offering pre-trained models for PASCAL VOC and INRIA Person datasets.

How It Works

The system employs mixtures of deformable part models, represented using a grammar formalism. It supports both latent SVM and weak-label structural SVM (WL-SSVM) for learning. The implementation includes features like a scale and location prior, star-cascade detection, and context rescoring using class-specific SVMs, aiming for improved accuracy and efficiency over prior work.

Quick Start & Requirements

• Install: Unpack the code, run compile in MATLAB to build MEX helper functions (may require editing compile.m for system-specific convolution routines).
• Prerequisites: Linux or OS X, MATLAB, GCC >= 4.2 (with OpenMP), PASCAL VOC devkit (download and configure voc_config.m).
• Resources: At least 4GB RAM (plus ~0.75GB per parallel MATLAB worker). Tested with MATLAB R2011a.
• Demo: Run demo.m or demo_cascade.m.
• Docs: FAQ available at http://people.cs.uchicago.edu/~rbg/latent/voc-release5-faq.html.

Highlighted Details

• Implements weak-label structural SVM (wl-ssvm) and person grammar model.
• Includes context rescoring and star-cascade detection algorithms.
• Supports multicore processing via MATLAB Parallel Computing Toolbox (matlabpool).
• Offers pre-trained models for PASCAL VOC 2007/2010 and INRIA Person datasets.

Maintenance & Community

This is a release from a Ph.D. dissertation, with the primary contributor being Ross Girshick. Further development or active community support is not explicitly indicated.

Licensing & Compatibility

The README does not explicitly state a license. It mentions support from the National Science Foundation. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The code is implemented in MATLAB and requires MEX compilation, potentially leading to system-specific setup challenges. The README notes that GitHub code may not be as thoroughly tested as the official tarball release. Configuration requires downloading and setting up the PASCAL VOC devkit.