Continuous Pose Estimation with a Spatial Ensemble of Fisher Regressors

Abstract

In this paper, we treat the problem of continuous pose estimation for object categories as a regression problem on the basis of only 2D training information. While regression is a natural framework for continuous problems, regression methods so far achieved inferior results with respect to 3D-based and 2D-based classification-and-refinement approaches. This may be attributed to their weakness to high intra-class variability as well as to noisy matching procedures and lack of geometrical constraints.We propose to apply regression to Fisher-encoded vectors computed from large cells by learning an array of Fisher regressors. Fisher encoding makes our algorithm flexible to variations in class appearance, while the array structure permits to indirectly introduce spatial context information in the approach. We formulate our problem as a MAP inference problem, where the likelihood function is composed of a generative term based on the prediction error generated by the ensemble of Fisher regressors as well as a discriminative term based on SVM classifiers.We test our algorithm on three publicly available datasets that envisage several difficulties, such as high intra-class variability, truncations, occlusions, and motion blur, obtaining state-of-the-art results.

Self-Supervised Learning for Specified Latent Representation

“…In Table II, we present the results from the literature and the results of our three approaches with the same network architecture (our approach 1, our approach 2, and our approach 3). As seen, our approach 1 obtains a relative improvement in the MeanAE of approximately 31.8% with respect to state of the art [40]. Our approach 2 uses only 123 labeled samples with the regression loss and 1389 structured unlabeled samples with the structure loss but still outperforms the state of the art [40] by 11.6% in MeanAE.…”

“…As seen, our approach 1 obtains a relative improvement in the MeanAE of approximately 31.8% with respect to state of the art [40]. Our approach 2 uses only 123 labeled samples with the regression loss and 1389 structured unlabeled samples with the structure loss but still outperforms the state of the art [40] by 11.6% in MeanAE. In addition, our approach 3 using only the 123 labeled samples with the regression loss is not as good as our approach 2, which can prove the effectiveness of the structure loss.…”

Estimating Small Differences in Car-Pose from Orbits

“…[16] uses probabilistic regression based on Hough Forests with an uncertainty criterion for continuous pose estimation. Similarly, [6] uses regressors on Fisher-encoded vectors extracted from spatial cells. [7] trains a separate SVM-classifier per viewpoint bin using off-the-shelf CNN-features.…”

Class-Specific Object Pose Estimation and Reconstruction Using 3D Part Geometry

“…The performance of statistical 2D feature based methods from the computer vision research literature inspired the development of most distinctive view-based techniques. Existing techniques [4,6,8] treat viewpoint estimation as a classification problem by dividing the viewpoint range into discrete bins. Ghodrati et al [6] train multiple Support Vector Machine (SVM) classifiers, one for each discrete viewpoint, treating each classifier independently of the others.…”

“…He et al [8] use a two-step process, wherein a viewpoint-parametrized classifier is first used to estimate a coarse viewpoint followed by fine-tuning step. Fenzi et al [4] treat continuous viewpoint estimation as a regression problem which is solved using a Radial Basis Function Neural Network (RBF-NN). The RBF-NN is trained to predict the appearance features as a function of the viewpoint.…”