Devansh Arpit

Person Re-identification for Improved Multi-person Multi-camera Tracking by Continuous Entity Association

We present a novel approach to person tracking within the context of entity association. In large-scale distributed multi-camera systems, person re-identification is a challenging computer vision task as the problem is two-fold: detecting entities through identification and recognition techniques; and connecting entities temporally by associating them in often crowded environments. Since tracking essentially involves linking detections, we can reformulate it purely as a re-identification task. The inherent advantage of such a reformulation lies in the ability of the tracking algorithm to effectively handle temporal discontinuities in multi-camera environments. To accomplish this, we model human appearance, face biometric and location constraints across cameras. We do not make restrictive assumptions such as number of people in a scene. Our approach is validated by using a simple and efficient inference algorithm. Results on two publicly available datasets, CamNeT and DukeMTMC, are significantly better compared to other existing methods.

Subspace learning via low rank projections for dimensionality reduction

Subspace learning algorithms aim at finding low dimensional linear manifolds that are representative of the data at hand. In this paper we propose a semi-supervised approach that fits any given dataset to a low dimensional subspace while maintaining class separability. Our approach has no tunable parameters as against many existing subspace learning algorithms which obviates the need for cross-validation. We apply our algorithm to the problem of face recognition. We perform both qualitative as well as quantitative experiments on multiple real world datasets. For qualitative analysis we visualize the class separability of binary and multi-class projected data. For quantitative analysis, we perform classification experiments on projected data and achieve state-of-the-art results compared to popular existing dimensionality reduction methods.