Multiclass Discriminative Fields for Parts-Based Object Detection

In this paper, we present a discriminative framework for parts-based object detection based on the multiclass extensions of binary discriminative fields described in [1]. These fields allow simultaneous discriminative modeling of the appearance of individual parts and the geometric relationship between them. The conventional Markov Random Field (MRF) formulations cannot be used for this purpose because they do not allow the use of data while modeling interaction between labels which is crucial for enforcing geometric consistencies between parts. The proposed technique can handle object deformations, occlusions and multiple-instance detection in a single trained model with no added computational efforts. The parameters of the field are learned using efficient maximum marginal approximations and inference is carried out using loopy belief propagation. We demonstrate the efficacy of this approach through controlled preliminary experiments on rigid and deformable synthetic toy objects.