Claudio Baecchi
Claudio Baecchi was born on December 14th, 1984 in Florence. He received a laurea degree in computer engineering from University of Florence, with a thesis on Fisher Feature Fusion Forests for visual object recognition. Currently he’s working at the Visual Information and Media Lab at Media Integration and Communication Centre, University of Florence.
Hi,
my name is Dino Pačandi and I’m trying to reproduce the image classification results that you reported in your article “Fisher vectors over random density forests for object recognition”, ICPR 2014.
The part that I find hard to reproduce is getting fully balanced trees using the proposed splitting method. To be more specific, when using forest with 4 trees and max tree depth set to 6, my Fisher Vector dimension is around 24k while yours is around 32k.
My question here – are there any details that you might have omitted from the article that can help achieve fully balanced trees?
Also, I’m interested in how did you acquire SIFT descriptors; library used (I’m using VLFeat), scales and steps.
Regards,
Dino Pačandi