• Video analysis
• Image/vector indexing
• Image classification
• Publications
• CV
• Teaching

Image classification projects

PQ-fication of a classification method

I extended the trace-norm regularized classification method by Zaid Harchaoui to handle PQ-compressed descriptor vectors, to improve scalability, see our CVPR 2012 paper. It consisted in optimizing the right and left multiplication with a PQ-compressed matrix. This can be done in the compressed domain but is hard to do faster than an optimized BLAS.

Object detection

With Navneet Dalal, I participated in the Pascal VOC 2006 detection track (which we won). I also participated in ImageCLEF 2009, we won the image-only modality (see this workshop paper).

I prepared for release the MashFishDet software by Gokberk Cinbis, that peforms object detection from image segmentation proposals.

I implemented the initial JSGD package for large-scale classifier training using stochastic gradient descent. It handles PQ compressed input vectors.

Deep learning

I developed a Python wrapper for the deep learning package Caffe in use at LEAR. The main objective was to make it more flexible, by not requiring protobuf configuration files, and to access the activations and weights with numpy arrays at any point of the training or testing.