Parameterizing Object Detectors in the Continuous Pose Space

Abstract

Abstract. Object detection and pose estimation are interdependent problems in computer vision. Many past works decouple these problems, either by discretizing the continuous pose and training pose-specific object detectors, or by building pose estimators on top of detector outputs. In this paper, we propose a structured kernel machine approach to treat object detection and pose estimation jointly in a mutually benificial way. In our formulation, a unified, continuously parameterized, discriminative appearance model is learned over the entire pose space. We propose a cascaded discrete-continuous algorithm for efficient inference, and give effective online constraint generation strategies for learning our model using structural SVMs. On three standard benchmarks, our method performs better than, or on par with, state-of-the-art methods in the combined task of object detection and pose estimation.

Pose Estimation of Object Categories in Videos Using Linear Programming

“…In particular, our Mean Absolut Error (MAE) is approximately 80% smaller with respect to the results published in [11], which is the most accurate method on this dataset so far. As the results in the first two columns show, the overall mean is strongly affected by few yet detrimental 180…”

“…We compare our method to six single-frame state-of-theart pose estimators [3,2,11,10,22,19] and one videobased approach that has been recently proposed [20] used the same testing framework as previous works, i.e., two different splits for training and testing. (i) 50% Split: training the model on the first 10 sequences and testing it on the remaining 10; (ii) Leave-One-Out: training the model on 19 sequences and testing it on the remaining one.…”

“…This is the case of autonomous driving and augmented reality applications, where the object pose must be accurately estimated. Most of the approaches proposed in the literature address the pose estimation problem of object classes for still images [19,3,10,11], while just a small number of We propose a method for pose estimation in video sequences that finds the globally optimum pose trajectory given the single-frame pose distributions learned using regression. We create a graph where the nodes are sampled from single-image pose distributions (black circles).…”

Driver head pose estimation using efficient descriptor fusion

“…In [26], normalized faces are used to train an auto-associative memory using the Widrow-Hoff correction rule in order to classify head poses. In one of the most recent works [27], the authors consider that object detection and continuous pose estimation are interdependent problems and they jointly formulate them as a structured prediction problem, by learning a single and continuously parameterized object appearance model over the entire pose space. After that, they design a cascaded discretecontinuous inference algorithm to effectively optimize a non-convex objective, by generating a diverse proposal to explore the complicated search space.…”

A Novel Representation of Parts for Accurate 3D Object Detection and Tracking in Monocular Images

“…However, each of the current approaches has its own weaknesses: Many of these approaches [5,13,1,35] rely on a depth sensor, which would fail on metallic objects or outdoor scenes; methods based on feature points [25,17] expect textured objects; those based on edges [4,39] are sensitive to cluttered background; most of these methods [13,27,30,38,11,45,20] are not robust to occlusion. We also want a method fast enough for interactive 3D applications.…”