Implementation for: Polceanu M., Harrouet F., Buche C., "Fast Multi-Scale fHOG Feature Extraction Using Histogram Downsampling", RoboCup International Symposium, 2018.
This code uses a modified version of Dlib 19.4 (http://dlib.net/) which is included in this repository.
The training and testing datasets are included. Images are obtained from multiple sources, including the SPQR dataset (http://www.dis.uniroma1.it/~labrococo/?q=node/459).
Level up to which scheme is applied is hard-coded to 7 in dlib-19.4-modified/dlib/image_processing/scan_fhog_pyramid.h line:928 (link to line) (sorry, research code...). After changes, the library must be recompiled.
Other parameters are found in the same file as above, at lines 685-693.
Configurations:
Original baseline:
- set USE_HACK to 0
Power law (Dollar et al. 2014):
- set USE_HACK to 1
- set hybridLevel to 10
Paper results:
- set USE_HACK to 1
- set hybridLevel to values between 2 and 9
Note: Average precision on the testing dataset will be lower than those reported in the paper. Results from the paper were obtained by doing a hyperparameter search at each level. This implementation has hardcoded hyperparameters and is only optimized for the original baseline.
git clone https://github.com/polceanum/fast.fhog.git
cd fhog
cd dlib-19.4-modified
mkdir build
cd build
cmake ..
make -j
cd ../..
mkdir build
cd build
cmake ..
make -j
./learnRobot ../dataset ../dataset
[...]
objective: 1398.83
objective gap: 1398.82
risk: 46.8166
risk gap: 46.8166
num planes: 3
iter: 1
objective: 506.993
objective gap: 506.687
risk: 16.9581
risk gap: 16.9581
num planes: 4
iter: 2
objective: 772.134
objective gap: 771.459
risk: 25.8197
risk gap: 25.8197
num planes: 5
iter: 3
[...]
hack time: 0.00600257
avg hack time: 0.00665809
correct_hits=161 total_hits=168 total_true_targets=185
testing results (precision, recall, average precision): 0.958333 0.87027 0.865549
[...]
Robot detection:
Learned vignette:
Extreme lighting conditions:
