Abstract: Transformer models have shown great success handling long-range interactions, making them a promising tool for modeling video. However, they lack inductive biases and scale quadratically with input length. These limitations are further exacerbated when dealing with the high dimensionality introduced by the temporal dimension. While there are surveys analyzing the advances of Transformers for vision, none focus on an in-depth analysis of video-specific designs. In this survey, we analyze the main contributions and trends of works leveraging Transformers to model video. Specifically, we delve into how videos are handled at the input level first. Then, we study the architectural changes made to deal with video more efficiently, reduce redundancy, re-introduce useful inductive biases, and capture long-term temporal dynamics. In addition, we provide an overview of different training regimes and explore effective self-supervised learning strategies for video. Finally, we conduct a performance comparison on the most common benchmark for Video Transformers (i.e., action classification), finding them to outperform 3D ConvNets even with less computational complexity.

Abstract: In this paper, we present a fully-automatic Spatio-Temporal GrabCut human segmentation methodology that combines tracking and segmentation. GrabCut initialization is performed by a HOG-based subject detection, face detection, and skin color model. Spatial information is included by Mean Shift clustering whereas temporal coherence is considered by the historical of Gaussian Mixture Models. Moreover, full face and pose recovery is obtained by combining human segmentation with Active Appearance Models and Conditional Random Fields. Results over public datasets and in a new Human Limb dataset show a robust segmentation and recovery of both face and pose using the presented methodology.

Abstract: This work proposes a fully convolutional network architecture for RGB face image generation from a given input thermal face image to be applied in face recognition scenarios. The proposed method is based on the FusionNet architecture and increases robustness against overfitting using dropout after bridge connections, randomised leaky ReLUs (RReLUs), and orthogonal regularization. Furthermore, we propose to use a decoding block with resize convolution instead of transposed convolution to improve final RGB face image generation. To validate our proposed network architecture, we train a face classifier and compare its face recognition rate on the reconstructed RGB images from the proposed architecture, to those when reconstructing images with the original FusionNet, as well as when using the original RGB images. As a result, we are introducing a new architecture which leads to a more accurate network.