Synthetic Humans for Action Recognition from Unseen Viewpoints

Although synthetic training data has been shown to be beneficial for tasks such as human pose estimation, its use for RGB human action recognition is relatively unexplored. Our goal in this work is to answer the question whether synthetic humans can improve the performance of human action recognition, with a particular focus on generalization to unseen viewpoints. We make use of the recent advances in monocular 3D human body reconstruction from real action sequences to automatically render synthetic training videos for the action labels. We make the following contributions: (1) we investigate the extent of variations and augmentations that are beneficial to improving performance at new viewpoints. We consider changes in body shape and clothing for individuals, as well as more action relevant augmentations such as non-uniform frame sampling, and interpolating between the motion of individuals performing the same action; (2) We introduce a new data generation methodology, SURREACT, that allows training of spatio-temporal CNNs for action classification; (3) We substantially improve the state-of-the-art action recognition performance on the NTU RGB+D and UESTC standard human action multi-view benchmarks; Finally, (4) we extend the augmentation approach to in-the-wild videos from a subset of the Kinetics dataset to investigate the case when only one-shot training data is available, and demonstrate improvements in this case as well.

The Existence of FGDRP$(3,g^u)'$s

By an FGDRP$(3,g^u)$, we mean a uniform frame $(X,\cal G,\cal A)$ of block size 3, index 2 and type $g^u$, where the blocks of $\cal{A}$ can be arranged into a $gu/3\times gu$ array. This array has the properties: (1) the main diagonal consists of $u$ empty subarrays of sizes $g/3\times g$; (2) the blocks in each column form a partial parallel class partitioning $X \setminus G$ for some $G\in \cal G$, while the blocks in each row contain every element of $X \setminus G$ $3$ times and no element of $G$ for some $G\in \cal{G}$. The obvious necessary conditions for the existence of an FGDRP$(3,g^u)$ are $u\geq 5$ and $g\equiv 0$ (mod 3). In this paper, we show that these conditions are also sufficient with the possible exceptions of $(g,u)\in \{(6,15),(9,18),(9,28),(9,34),(30,15)\}$.

Zoom Rendering: Improving 3-D Rendering Performance with 2-D Operations

We propose a new method to accelerate rendering during the interactive visualization of complex scenes. The method is applicable if the cost of per-pixel processing is high compared to simple frame buffer transfer operations, as supported by low-end graphics systems with high-performance CPUs or 2-D graphics accelerators. In this case rendering an appropriately down-scaled image and then enlarging it allows a trade-off of rendering speed and image quality. Using this method, more uniform frame rates can be achieved and the dynamic viewing error can be reduced.

Samuel compactification and completion of uniform frames

The aim of this paper is twofold: first, to construct the compact regular coreflection of uniform frames, that is, the frame counterpart of the Samuel compactification of uniform spaces (Samuel [10]), and then to use this for a new description of the completion of a uniform frame, as an alternative to those previously given by Isbell [6] on the one hand and Kříž [8] on the other. In addition, we present a few further results, as well as new proofs of known ones, that are naturally connected with completions and arise particularly easily from our approach to them. Most prominently among these, we identify the uniform space of minimal Cauchy filters of a uniform frame as the spectrum of its completion.

Torsional provisions in building codes

A study is made of the accuracy of the static code provisions on torsional effects, with special reference to the National Building Code of Canada of 1977. A uniform frame type monosymmetric 12-story building is used as an example. The static story torque is compared with the dynamic torque computed using the response spectrum technique as outlined in Commentary K of the Code. It has been found that for a building with uniform eccentricity, the static code torque estimate is good if the effect of sympathetic coupled torsional–lateral resonance is small. At the sympathetic coupled resonance, the static code torsional provision underestimates the story torque by a factor of approximately two. Also, it is shown that for buildings with large eccentricities, sympathetic resonance is unlikely to occur and the current Code requirement of doubling the computed torque for design is a very conservative requirement.In addition, the eccentricity expression given in the 1980 Code is used to estimate the story torques for eccentrically set-back buildings and its accuracy is checked with a response spectrum technique. It is concluded that for irregular structures, such as eccentrically set-back buildings, a dynamic analysis is the only reliable procedure for the estimation of the torque loading on the building.