The original paper refers to:
Integral Channel Features, P. Dollar, Z. Tu, P. Perona, and S. Belongie, BMVC 2009
The improved version refers to:
Pedestrian Detection at 100 Frames per Second, R. Benenson, M. Mathias, R. Timofte, and L. Van Gool, CVPR 2012
Seeking the Strongest Rigid Detector, R. Benenson, M. Mathias, R. Timofte, and L. Van Gool, CVPR 2013
This is a long story, you should read the original paper and the two follow ups to get the idea why ICF is the strongest rigid detector, ccv does this though:
./icfdetect <Your Image contains Pedestrians> ../samples/pedestrian.icf | ./icfdraw.rb <Your Image contains Pedestrians> output.png
Checkout the output.png, all pedestrians should have a red box on them.
ICF has two modes, one is presented on the original paper, by resizing input into different scales, and then run the same classifier again and again on these resized inputs. The second is presented in the improved version, by running multiple classifiers that are trained on different scales on the same input.
The second approach will be the faster alternative, unfortunately, I am unable to obtain a reasonable recall / precision with the second approach.
Running in the first mode, on a computer with Core i7 3770K, with INRIA 2008 test set, the figures are:
real 2m19.18s user 2m16.30s sys 0m2.79s
It is still slower than HOG, but faster than DPM implementation in libccv.
The pedestrian.icf model provided in ./samples are trained with INRIA 2008 training dataset, but with additional 7542 negative samples collected from VOC2011. The model is trained at size 31x74, with 6px margins on each side.
The provided model is then tested with INRIA 2008 test dataset, if bounding boxes overlap is greater than 0.5 of the bigger bounding boxes, it is a true positive. The validation script is available at ./bin/icfvldtr.rb.
Which has roughly the same recall as DPM implementation provided in ccv, with roughly the same false alarms too.
ccv provides utilities to train your own object models. Specifically, for ICF, these utilities are available at ./bin/icfcreate and ./bin/icfoptimize.
Will show you the parameters that ccv supports when training an object model.
If you have libdispatch installed and properly enabled on your machine, ccv will utilize all your CPU cores to speed up the training process.
The INRIA pedestrian dataset can be downloaded from:
The annotation format is substantially different from what ccv requires, I use this simple script to extract annotations from INRIA dataset:
You also want to have a collection of background (none pedestrian) files, I combined data from both INRIA and VOC2011 to generates that list:
find ../data/negs/*.jpg > no-pedestrian.txt
After all these ready, and have a PC with enough computational power:
./icfcreate --positive-list pedestrian.icf_samples --background-list no-pedestrian.txt --validate-list pedestrian.icf_test --negative-count 10000 --positive-count 10000 --feature-size 50000 --weak-classifier-count 2000 --size 30x90 --margin 10,10,10,10 --working-dir icf-data --acceptance 0.7 --base-dir ../data/INRIAPerson/Train/pos/
The classifier cascade will be bootstrapping 3 times, pooling from 50,000 features, and the final boosted classifier will have 2,000 weak classifier. On the PC that I am running (with SSD / hard-drive hybrid (through flashcache), 32GiB memory and Core i7 3770K), it takes a day to finish training one classifier. At minimal, you should have about 16GB available memory to get the program finish running.
The final-cascade file in your working directory is the classifier model file that you can use. Using ./bin/icfoptimize, you should be able to set proper soft cascading thresholds for the classifier to speed up detection:
comments powered by Disqus
./icfoptimize --positive-list pedestrian.icf_test --classifier-cascade icf-data/final-cascade --acceptance 0.7 --base-dir ../data/INRIAPerson/Test/pos/