3D Bird Reconstruction: A Dataset, Model, and Shape Recovery from a Single View

Progress on Computational Ethology

“…但是其中大部分是一些方法学的研究, 还缺乏与 脑中行为控制机制的关联分析. 从2015年起, 计算机视 觉领域开始关注利用深度学习解决从二维物体图像估 计三维形状的问题 [50] , 并在鸟类 [51] 、犬 [52] 以及其他多 种动物 . 数据降维有主成分分析(principal component analysis) [62] 、t分布随机邻近嵌入法(t-distributed stochastic neighbor embedding) [58] 、UMAP(uniform manifold approximation and projection for dimension reduction) [43] 等方法.…”

Birds of a Feather: Capturing Avian Shape Models from Images

“…Biggs et al [8] extend the SMAL model to include limb scaling to model dog breeds. Badger et al [7] similarly parametrized an articulated bird mesh and used multi-view data from an aviary to provide pose and shape priors.…”

“…Changes in shape can convey an animal's health status and transmit social signals [1,34]. Recent methods that use articulated mesh models to extract these signals from images [7,8,54,55] are poised to transform these fields. One challenge that prevents wider adoption of this approach, however, is the difficulty of obtaining suitable models for new species.…”

“…Recent articulated, 3D animal shape models obtain training data from 3D scans of toy animals figurines, or multiple views of the same subject [7,56,55]. They produce great quality of reconstruction when the target is represented in the source data.…”

“…We propose a method to directly capture articulated shapes from species-specific image collections, when a strong deformable shape prior is not available. We focus our effort on CUB and start with an articulated bird mesh [7] as a generic template model. For a given collection, we first align the mesh to each annotated instance by solving for the pose and camera parameters.…”