Version 1.1 (9 May 2016)

Contributed by Bumsub Ham (bumsub.ham@inria.fr) and Minsu Cho (minsu.cho@inria.fr).

This code is written in MATLAB, and implements the ProposalFlow and its benchmark in [1]. For the PF dataset, see our project page: http://www.di.ens.fr/willow/research/proposalflow.

We use the PF dataset (included) to evaluate sparse and dense versions of ProposalFlow.

• Download [VLFeat] (http://www.vlfeat.org/) and [MatConvNet] (http://www.vlfeat.org/matconvnet/).
• Download the following source codes of object proposal or other proposal methods you would like to test:
  • [EdgeBox] (https://github.com/pdollar/edges);
  • [SelectiveSearch] (http://koen.me/research/selectivesearch/);
  • [Randomized Prim’s] (https://github.com/smanenfr/rp#rp);
  • [Multiscale Combinatorial Grouping] (https://github.com/jponttuset/mcg);
• Download a ImageNet [Caffe Reference model] (http://www.vlfeat.org/matconvnet/pretrained/) in ./feature/cnn-model/.

Set the file path of these libraries in set_path.m and matching configulartion (object class, types and numbers of object proposals, and feature) in set_conf_WILLOW.m in ./PF-dataset-WILLOW-code/, and run

demo_BM_PF_WILLOW.m 

If you just want to compute dense flow fields such as SIFTFlow [2], run

./_demo-DenseFlow/demo_DenseFlow.m 

• prepKP_WILLOW.m: load keypoint annotations and save them as a file.
• ext_proposal_WILLOW.m: extract object proposals from images.
• ext_active_proposal_WILLOW.m: extract valid object proposals (object proposals near object bounding boxes).
• makeGT_WILLOW.m: automatically estimate ground-truth matches for valid object proposals using the keypoint annotations and TPS warping.
• ext_feature_WILLOW.m: extract feature descriptors for all object proposals.
• matching_WILLOW.m: compute proposal flow (matching all object proposals between two images).
• eva_WILLOW.m: evaluate the PCR and mIoU@k performance of proposal flow.
• eva_avg_WILLOW.m: evaluate proposal flow (averaging performance per feature).
• dense_flow_WILLOW.m: compute dense flow fields using proposal flow.
• dense_flow_eva_WILLOW.m: evaluating dense flow field (PCK performance).

• do_readKP_WILLOW.m: visualize annotations.

• The code is provided for academic use only. Use of the code in any commercial or industrial related activities is prohibited.
• If you use our code or dataset, please cite the paper.

@InProceedings{ham2016, author = {Bumsub Ham and Minsu Cho and and Cordelia Schmid and Jean Ponce}, title = {Proposal Flow}, booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE}, year = {2016} } 

• This code uses the author provided source codes for generating object proposals: SelectiveSearch, Randomized Prim’s, EdgeBox, Multiscale Combinatorial Grouping, [Sliding Window, Uniform Sampling, and Gaussian Sampling] (https://github.com/hosang/detection-proposals).

• For CNN features, this code uses a ImageNet Caffe Reference model: AlexNet trained on ILSVRC 2012, with a minor variation from the version as described in ImageNet classification with deep convolutional neural networks by Krizhevsky et al. in NIPS 2012.

• Version 1.0 (28 Mar 2016)
  • Inirial release
• Version 1.1 (9 May 2016)
  • Improved matching speed (LOM.m).

[1] B. Ham, M. Cho, C. Schmid, and J. Ponce, "Proposal Flow", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.

[2] C. Liu, J. Yuen, and A. Torralba, "Sift flow: Dense correspondence across scenes and its applications", IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2011.