Pose Estimation Diagnostic Tool

INTRODUCTION:

This project contains the source code and annotations for analyzing object detectors and pose estimators on the PASCAL 3D+ dataset.

This is a repository with an implementation of the diagnostic tool described in our ECCV2016 paper. We provide here the codes and data needed to reproduce all the experiments detailed in the paper.

LICENSE

The license information of this project is described in the file "LICENSE".

CITING

If you make use of this software, please cite the following reference in any publications:

    @inproceedings{Redondo-Cabrera2016,
           Title  = {Pose Estimation Errors, the Ultimate Diagnosis},
           Author = {Redondo-Cabrera, C. and Lopez-Sastre, R.~J. and Xiang, Y. and Tuytelaars, T. and Savarese, S.},
           Booktitle = {ECCV},
           Year = {2016}
    }

REQUIREMENTS

The diagnostic tool is developed and tested under Ubuntu 14.04. Matlab is required. To generate the PDF with the report, we use the tool pdflatex, which has to be correctly installed.

HOW TO GENERATE THE REPORTS?:

The tool we provide here generates the detailed reports described in our paper, for any method using the PASCAL 3D+ dataset.

• CASE I: How to generate the reports for the methods VDPM (vdpm), V&K (vpskps), DPM+VOC-VP (3ddpm) or BHF (bhf) pose estimators? Note these are the methods analyzed in the paper.

  1. Download the PASCAL 3D+ dataset (Release 1.1) and uncompress the zip file provided in the folder PASCAL3D+.

  2. In the script src/poseEstimationAnalysisScript.m, set to 1 all flags (SKIP_SAVED_FILES, SAVE_QUALITATIVE, SHOW_FIGURES, DO_TEX, DO_OVERLAP_CRITERIA_ANALYSIS and SAVE_SUMMARY).

  3. In the script src/poseEstimationAnalysisScript.m, add to the ''detectors'' list the ones you want to be diagnosed. By default, the tool generates the report for a random pose assignment named 'rand-gt'.

  4. Open Matlab, go to the src folder and run the script poseEstimationAnalysisScript.m. The PDF with the detailed report will be generated in the path: results/method_name/tex/AnalysisAutoReportTemplate.pdf

• CASE II: How to generate the reports for your own method using the PASCAL 3D+ dataset?

  1. Create a subdirectory in the ''detections'' folder. Assign this directory the name of your new model.

  2. In this directory create a text file for each object category. These files must contain in each row the detections and pose estimations obtained by the model, following this format:

        image_name (without extension, \ie '.jpg') detector_score x1 y1 x2 y2 azimuth zenith
        example: 2008_000002 0.292526 34.00 11.00 448.00 293.00 342.86 171.43                 
     

  3. Add a corresponding entry to the script src/setDetectorInfo.m and update the ''detectors'' variable in src/poseEstimationAnalysisScript.m to analyze the created model.

  4. Perform steps [2-4] of CASE I.