Research – recent

I am mostly interested in images and in their mathematical representation and interpretation. My work covers diverse topics such as statistical analysis of visual tasks, segmentation and recognition algorithms, learning, and recently, also 3D point cloud analysis. A selection of my projects over the years is described in the Research page.

My recent projects are described here. The whole research field was transformed in the last 10 year with the introduction of the powerful deep learning algorithms. While I am not fully converted, part of my work goes in this direction as well.

Recognition under occlusion

Objects in real images are often partially occluded, and deep categorization algorithms usually fail for them. The usual approach is to train on occluded object, but we have shown that keeping the filters narrow, by specialized regularization, is as good.

rIncreasing cnn robustness to occlusions by reducing filter suppot, E Osherov, M Lindenbaum, CVPR, 550-561, 2017

Occlusion

Generic (non-semantic) segmentation

Learning region representation (with an example to its embedding shown on the right) and adding it to edge data improved the state the art in generic edge detection.

O Isaacs, O Shayer, M Lindenbaum, CVPR 2020.
BRRF-rep-example

Recognition from partial minimal data

Can we recognize an object from a single small patch? The human visual systems can do it. Here we proceed a computational mechanism that can do it as well with similar characteristics. (In the example the green patch is of the minimal size.)

On the Minimal Recognizable Image Patch, M Fonaryov, M Lindenbaum, ICPR, 6734-6741, 2021.

Screen Shot 2022-03-02 at 20.04.24

Strcture from motion with a-contrario

Looking at randomly placed points, the remarkable human visual system can detect a subset of points that coherently move. We analyze this (well known)  phenomena and provide quantitative predictions regarding the visibility using the statistical a contrario methodology. 

Seeing Things in Random-Dot Videos, T Dagès, M Lindenbaum, AM Bruckstein, ACPR, 195-208, 2020.

3D point cloud data

The 3D modified Fisher vector (3DmFV) representation

Representing a point cloud in a structures way is challenging. We proposed a method that is based on the classical Fisher vector representation of a data set with respect to a distribution, and a deep neural algorithm.

3dmfv: Three-dimensional point cloud classification in real-time using convolutional neural networks, Y Ben-Shabat, M Lindenbaum, A Fischer, IEEE Robotics and Automation Letters 3 (4), 3145-3152, 2018.

3Ddata

The Nesti-Net surface normal estimator

Estimators for the surface normal and the  effective scale,  that is based on the 3DmFV. 

Nesti-net: Normal estimation for unstructured 3d point clouds using convolutional neural networks, Y Ben-Shabat, M Lindenbaum, A Fischer, CVPR, 2019.

image 2

Comparing point clouds

An accurate estimator for comparing two point clouds, using the estimated distance from a point to the underlying point cloud surface that is based on local (and hence fast) modeling  

DPDist: Comparing point clouds using deep point cloud distance, D Urbach, Y Ben-Shabat, M Lindenbaum, ECCV, 545-560, 2020.

DPdist