Diagnosing Object Detectors

Citation

This project is modified from here. Please cite it in your publications if it helps your research:

@article{hoiem2012diagnosing,
  title={Diagnosing error in object detectors},
  author={Hoiem, Derek and Chodpathumwan, Yodsawalai and Dai, Qieyun},
  journal={Computer Vision--ECCV 2012},
  pages={340--353},
  year={2012},
  publisher={Springer}
}

Derek Hoiem July 30, 2012 (updated July 31, 2014)

This project contains the source code and annotations for analyzing object detectors with the PASCAL VOC 2007 dataset. The annotations were created by Yodsawalai Chodpathumwan in Spring 2010 and Fall 2011 as an undergraduate RA. The associated published paper is:

D. Hoiem, Y. Chodpathumwan, and Q. Dai, Diagnosing Error in Object Detectors, ECCV 2012.

How to Run

CASE I: PASCAL VOC 2007, felzenszwalb or vedaldi detectors

1. In detectionAnalysisScript.m, set all flags on top to 1 (true) (note: DO_SHOW_SURPRISING_MISSES is optional)
2. Set the imdir path to a valid directory of VOC images
3. Run detectionAnalysisScript.m in Matlab

CASE II: PASCAL VOC 2007, your own detectors

1. Create a text file of detector outfits with rows of file_id conf x1 y1 x2 y2 (standard format, see readDetections.m). Put this file in a subdirectory within detections.
2. Add a corresponding entry to setDetectorInfo.m and update detector variable in detectionAnalysisScript.m.
3. Perform all steps of CASE I.

CASE III: later versions of VOC Annotations are not currently available for detailed analysis of true detections.

1. Modify setDatasetParameters.m:
   • Set the imdir, VOCsourcepath, and VOCset to correspond to the dataset
   • Set objnames_extra = {}; change objnames_extra in DISPLAY_TP to objnames_selected.
2. Run as in CASE I or II

Case IV: other datasets If detections are bounding boxes and evaluation criteria is similar, use 'voc_compatible' as detector type and update setDatasetParameters.m. Otherwise, more modification may be required; see details in Code_Explanation.txt.

Creat A Report

1. The tex file for a report will be created for you in results/detname/tex.
2. Automatically remove whitespace from the figures using an Adobe Acrobat batch process or this script for pdfcrop from Ross Girshick:

      #!/bin/sh
      find . -name "*.pdf" -print0 | xargs -P6 -0 -I file pdfcrop file file
   

3. Use a latex compiler on detectionAnalysisAutoReportTemplate to make the pdf.

Description of Folders

annotations: Contains annotations for objects (excluding "difficult" objects, as defined by VOC annotations). The file Labels.txt helps to explain the format. Each file also contains whether the object was detected by the Felzenszwalb or the Vedaldi detector and the corresponding confidence, but this need not be used.

src: Contains code for reading the annotations and storing them in the PASCAL VOC record structure. Also contains code for computing normalized average precision and making comparisons between different subsets of the data.

results: For each detector, contains tables of normalized AP for various subsets of data and comparison plots.

detections: For each detector, contains a list of all detections above some threshold for each object.

Description of Script/Functions

detectionAnalysisScript.m

• TP_ANALYSIS: correlates detections with ground truth, assigns attributes such as occlusion and aspect to each object, and computes performance measures for various subsets of objects
• FP_ANALYSIS: computes fraction of top false positives due to localization error, confusion with similar objects, etc.; also computes the AP impact of false positives
• TP_DISPLAY: creates plots showing performance of detector for different subsets of objects, a summary plot, a text summary of characteristics of missed objects, and images of objects that are less confidently detected than expected
• FP_DISPLAY: displays statistics of the frequency and impact of false positives, including creation of a table in a text file
• DO_TEX: creates tex files for compiling a report

Other useful functions

• displayTopFP: displays most confident false positives
• displayRankedPositives: shows detection confidences of a subset of objects, from most to least confident

Further Help

Contact dhoiem@illinois.edu if there is a problem with the annotations or code. If instructions are unclear or incomplete, please try to figure it out first and send me a corrected version of the instructions. If you can't figure it out, email me, and I will help as time allows.