Training deformable part models with decorrelated features: Supplementary material
نویسندگان
چکیده
An interesting question when training a DPM is: how many latent update iterations are required to saturate detection performance? To answer this question, we look at mAP plotted against latent update iterations. To make the experimental protocol more precise, we briefly review the DPM training procedure. DPM training proceeds through three distinct phases. First, single-component, root-only models are trained independently on disjoint subsets of the positive examples (split by aspect ratio and left vs. right facing instances). These root-only models are trained using LSVM or LLDA with the root’s position and scale regarded as latent. Second, the root-only models are combined into a single mixture model, but still without parts. This model is trained using LSVM or LLDA with the choice of component filter, position and scale treated as latent. Finally, parts are added to the model, and the whole model is trained with LSVM or (LM-)LLDA where the component choice, root filter position and scale, and part locations are all latent. Our experiments start from the beginning of the final phase, i.e. the point at which parts are added to the model. To speed up training, we use 200 negative images per classes as was justified in the main paper. We consider two methods for adding parts. The first method is the default in the voc-release5 source code: parts are added by covering high energy areas of the root filter and then upsampling the root filter weights to twice their original resolution. The second method uses the exact same configuration of parts as the first method, but rather than 1 2 4 8 20 22 24 26 28 30 32 34
منابع مشابه
Augmenting Deformable Part Models with Irregular-shaped Object Patches Supplementary Material
متن کامل
Analysis and Synthesis of Facial Expressions by Feature-Points Tracking and Deformable Model
Face expression recognition is useful for designing new interactive devices offering the possibility of new ways for human to interact with computer systems. In this paper we develop a facial expressions analysis and synthesis system. The analysis part of the system is based on the facial features extracted from facial feature points (FFP) in frontal image sequences. Selected facial feature poi...
متن کاملFrom Virtual to Reality: Fast Adaptation of Virtual Object Detectors to Real Domains
The most successful 2D object detection methods require a large number of images annotated with object bounding boxes to be collected for training. We present an alternative approach that trains on virtual data rendered from 3D models, avoiding the need for manual labeling. Growing demand for virtual reality applications is quickly bringing about an abundance of available 3D models for a large ...
متن کاملDeformable part models
This paper describes a discriminatively trained, multiscale, deformable part model for object detection. Our system achieves a two-fold improvement in average precision over the best performance in the 2006 PASCAL person detection challenge. It also outperforms the best results in the 2007 challenge in ten out of twenty categories. The system relies heavily on deformable parts. While deformable...
متن کاملTheoretical Formulations for Finite Element Models of Functionally Graded Beams with Piezoelectric Layers
In this paper an overview of functionally graded materials and constitutive relations of electro elasticity for three-dimensional deformable solids is presented, and governing equations of the Bernoulli–Euler and Timoshenko beam theories which account for through-thickness power-law variation of a two-constituent material and piezoelectric layers are developed using the principle of virtua...
متن کامل