Diagnosing Error in Object Detectors

Derek Hoiem and Qieyun Dai and Yodsawalai Chodpathumwan
Computer Vision Group
Department of Computer Science
University of Illinois at Urbana-Champaign

Overview

This work provides a set of tools for analyzing object detector performance.

Note: (11/12/14) The summary plots (e.g., "animal" or "vehicle") for displayDetectionTrend were computed incorrectly. The revised code is now in the .tar.gz file, but the pdfs have not been updated. Thanks to Shaoqing Ren for noticing the bug and providing the fix. Another method displayDetectionTrend2.m is also provided, which averages across tic marks to summarize several categories.

Note: (11/7/17) Somehow VOC2007_records_test.mat became corrupted, so I replaced it with a file generated by Hongyu Pang. Thanks to Hongyu for that.

Downloads

The following resources are available:

An updated version (v2) of the code/annotations: [src/data (84.5MB)]
Description of updates: [pdf]
Examples of automatic analysis reports: [dpm_v4.pdf] [vedaldi2009.pdf] [cnn7_bb.pdf]
Original version (v1) of the code/annotations (in case you have trouble with the new version): [src/data (69MB)]

Publications

Diagnosing Error in Object Detectors
Derek Hoiem, Yodsawalai Chodpathumwan, and Qieyun Dai
ECCV, 2012. [pdf] [slides]

This research is supported by NSF IIS-0904209: Physically Grounded Object Recognition, NSF IIS-1053668: CAREER award, and ONR MURI Grant N000141010934.

Home