No description, website, or topics provided.
Matlab C M C++
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
feat_cache/voc_2007_flip_0_layer_fc7_finetuned_0/positives hos: initial commit Jun 18, 2014
minFunc_2012
nms
projsplx hos: initial commit Jun 18, 2014
saved_foreground_coords
utils hos: initial commit Jun 18, 2014
vis
README.md
SVM_Cost_L1hinge.m
SVM_Cost_Logistic.m
SVM_Cost_SmoothHinge.m Added missing SVM loss functions Jul 25, 2014
cluster_patches_parallel_single_nogt_20x1.m hos Jun 18, 2014
eval_classes_20x1_smooth_lsvm_topK_bagmine_greedycover_new.m hos Jun 18, 2014
feat_stats_hos.m
general_smooth_hinge.m hos Jun 18, 2014
load_cached_features_hos.m hos Jun 22, 2014
parse_smooth_lsvm_topK_bagmine_greedycover_maxiter_june17.m hos Jun 18, 2014
slslvm_cost_smoothhinge_bagmine.m
test_classes_hos_append_filename.m
train_classes_20x1_smooth_greedycover.m
train_classes_20x1_smooth_lsvm_topK_bagmine_greedycover.m hos Jun 18, 2014
voc_config.m
xform_feat_custom.m hos added missing xform_feat_custom.m Jul 25, 2014

README.md

Song-ICML2014

Citing the paper

@inproceedings{Song-ICML2014,
	title = "On learning to localize objects with minimum supervision",
	booktitle = "International Conference on Machine Learning (ICML)",
	year = "2014", 
	author = "Hyun Oh Song and Ross Girshick and Stefanie Jegelka and Julien Mairal and Zaid Harchaoui and Trevor Darrell",
}

License

This code is released under the Simplified BSD License (refer to the LICENSE file for details).

System Requirements

  • Linux
  • MATLAB

Install instructions

  1. Download and install R-CNN code (https://github.com/rbgirshick/rcnn)
  2. Following the instructions in R-CNN code, precompute fc7 features on the dataset (i.e. PASCAL VOC 2007)
  3. Unpack this code.
  4. Start Matlab where the unpacked code is.
  5. Mex compile the simplex projection cc code. >> curdir=pwd; cd('projsplx'); mex projsplx_c_float.cc; cd(curdir);

Train & Test work flow

Extract (not finetuned) and save fc7 features on all Selective Search windows on your dataset (i.e. PASCAL VOC2007) -> Cluster features -> Initial detector training via coverage maximization -> Refine via Smooth Latent SVM -> Test the detectors.

Usage

[A. Feature extraction on SS windows]

  1. Following the instructions in R-CNN code, precompute fc7 features.
  2. Save features under feat_cache/voc_2007_flip_0_layer_fc7_finetuned_0/

[B. Clustering] This is the slowest and the most computationally expensive step. Suggestions are welcomed. If you want to short cut and get a list of estimated foreground boxes on PASCAL 2007 trainval for all 20 classes. Check saved_foreground_coords directory.

  1. On cluster, launch cluster_patches_parallel_single_nogt_20x1.m in parallel with inputs (positive image id, classid). The first argument goes from 1 ~ #positive image per class. The second argument goes from 1 ~ 20.
  2. This will save about 70GB of clustering results for PASCAL VOC2007 under directory paris_results_nogt_20x1/%s_%d.mat where the first string is for the image id and the second integer if for class id (1~20).

[C. Initial detector training via coverage maximization]

  1. Start matlab.
  2. Launch train_classes_20x1_smooth_greedycover.m with input (classid)

[D. Refine via Smooth Latent SVM]

  1. Start matlab
  2. Launch train_classes_20x1_smooth_lsvm_topK_bagmine_greedycover.m with input (classid)

[E. Test the detectors]

  1. Start matlab
  2. Launch eval_classes_20x1_smooth_lsvm_topK_bagmine_greedycover.m with input (classid)
  3. (optional) to save a nicely parsed table of AP numbers for each class, launch parse_smooth_lsvm_topK_bagmine_greedycover_maxiter_june17.m.