In this work, we explore a variety of techniques and applications for visual problems involving videos of humans in the contexts of activity detection, pose detection, and forgery detection.

The first works discussed here address the issue of human activity detection in untrimmed video where the actions performed are spatially and temporally sparse. The video may therefore contain long sequences of frames where no actions occur, and the actions that do occur will often only comprise a very small percentage of the pixels on the screen. We address this with a two-stage architecture that first creates many coarse proposals with high recall, and then classifies and refines them to create temporally accurate activity proposals. We present two methods that follow this high-level paradigm: TRI-3D and CHUNK-3D.

This work on activity detection is then extended to include results on few-shot learning. In this domain, a system must learn to perform detection given only an extremely limited set of training examples. We propose a method we call a Self-Denoising Neural Network (SDNN), which takes inspiration from Denoising Autoencoders, in order to solve this problem, both in the context of activity detection and image classification.

We also propose a method that performs optical character recognition on real world images when no labels are available in the language we wish to transcribe. Specifically, we build an accurate transcription system for Hebrew street name signs when no labeled training data is available. In order to do this, we divide the problem into two components and address each separately: content, which refers to the characters and language structure, and style, which refers to the domain of the images (for example, real or synthetic). We train with simple synthetic Hebrew street signs to address the content components, and with labeled French street signs to address the style.

We continue our analysis by proposing a method for automatic detection of facial forgeries in videos and images. This work approaches the problem of facial forgery detection by breaking the face into multiple regions and training separate classifiers for each part. The end result is a collection of high-quality facial forgery detectors that are both accurate and explainable. We exploit this explainability by providing extensive empirical analysis of our method's results.

Next, we present work that focuses on multi-camera, multi-person 3D human pose estimation from video. To address this problem, we aggregate the outputs of a 2D human pose detector across cameras and actors using a novel factor graph formulation, which we optimize using the loopy belief propagation algorithm. In particular, our factor graph introduces a temporal smoothing term to create smooth transitions between poses across frames.

Finally, our last proposed method covers activity detection, pose detection, and tracking in the game of Ping Pong, where we present a new dataset, dubbed SPIN, with extensive annotations. We introduce several tasks with this dataset, including the task of predicting the future actions of players and tracking ball movements. To evaluate our performance on these tasks, we present a novel recurrent gated CNN architecture.