What makes for effective detection proposals?


Current top performing object detectors employ detection proposals to guide the search for objects, thereby avoiding exhaustive sliding window search across images. Despite the popularity and widespread use of detection proposals, it is unclear which trade-offs are made when using them during object detection. We provide an in-depth analysis of twelve proposal methods along with four baselines regarding proposal repeatability, ground truth annotation recall on PASCAL, ImageNet, and MS COCO, and their impact on DPM, R-CNN, and Fast R-CNN detection performance. Our analysis shows that for object detection improving proposal localisation accuracy is as important as improving recall. We introduce a novel metric, the average recall (AR), which rewards both high recall and good localisation and correlates surprisingly well with detection performance. Our findings show common strengths and weaknesses of existing methods, and provide insights and metrics for selecting and tuning proposal methods.


Please cite our paper, when you're using our data. :)

  • J. Hosang, R. Benenson, P. Dollár, and B. Schiele. What makes for effective detection proposals? PAMI 2015.
      author = {J. Hosang and R. Benenson and P. Doll\'ar and B. Schiele},
      title = {What makes for effective detection proposals?},
      journal = {PAMI},
      year = {2015}
  • J. Hosang, R. Benenson, and B. Schiele. How good are detection proposals, really? BMVC 2014.
    PDF, arXiv
      author = {J. Hosang and R. Benenson and B. Schiele},
      title = {How good are detection proposals, really?},
      booktitle = {BMVC},
      year = {2014}
For three different detection proposal methods, the picture show the four best localized proposals, i.e. closest to the ground truth. The methods are: Selective Search (red), BING (blue), Random Gaussian (green).

Data & Code

If you're interested in benchmarking your proposal detection method or in the code or data please contact me: Jan Hosang

I'm in the process of preparing data and code for release. You can find everything that is available so far below. Please let me know if you're missing something!

  • You can find the code at GitHub
  • For the recall experiments (see Section 4 in the paper), we run twelve detection proposal methods and four baseline methods over the pascal test set.

    • You can download the results of matching the proposals to the ground truth, which is smaller than all proposals: download (38MB, md5sum: aa574be08b8b235486b4e7cb49193ec9)
    • You can download the detection propsals: download (9.5GB, md5sum: 7a9ef3fa031e33bd044fc5ce4f1b83cf)

  • We also do recall experiments on ImageNet (see Section 4 in the paper).

    • Results of matching proposals to the ground truth, which is smaller than the set of all proposals: download (64MB, md5sum: fa6c9ea6c8bf0cb86d02b712aec6864d)
    • You can also download the proposals: download (9.7GB, md5sum: b67955b771f1304e70bd2cfd278d2de5)

  • Data from the recall experiments on COCO (see Section 4 in the paper).

    • Results of matching proposals to the ground truth: download (254MB, md5sum: 2e17c9998b3eaf2bff0a6b40916867bf)
    • Proposals: download (77GB, md5sum: efff6194c7597cccc4058e6103ccb59c)

  • For repeatability experiments (see Section 3 in the paper), we applied a number of perturbations to the pascal test set.

    • Here, you can download the transformed pascal images (70GB, md5sum: 5a3095a3ed893e098717fa9c88e19929)
    • You can download the result of the matching of proposals on the transformed pascal images to reproduce the plots (133MB, md5sum: 28a1722caf9e37b011d8b52ea24a7220)

Summary of methods and results

Average recall for different numbers of proposals

Figure 1: Average recall on the PASCAL VOC 2007 test set.
Figure 2: Average recall on the ImageNet 2013 validation set.
Figure 3: Average recall on the COCO 2014 validation set.

Detection performance with LM-LLDA and R-CNN on PASCAL VOC 2007