At Facebook, machine learning provides a wide range of capabilities that drive many aspects of user experience including ranking posts, content understanding, object detection and tracking for augmented and virtual reality, speech and text translations. While machine learning models are currently trained on customized data-center infrastructure, Facebook is working to bring machine learning inference to the edge. By doing so, user experience is improved with reduced latency (inference time) and becomes less dependent on network connectivity. Furthermore, this also enables many more applications of deep learning with important features only made available at the edge. This paper takes a data-driven approach to present the opportunities and design challenges faced by Facebook in order to enable machine learning inference locally on smart phones and other edge platforms.

Vision-based interfaces, such as those made popular by the Microsoft Kinect, suffer from the Midas Touch problem: every user motion can be interpreted as an interaction. In response, we developed an algorithm that combines facial features, body pose and motion to approximate a user's intention to interact with the system. We show how this can be used to determine when to pay attention to a user's actions and when to ignore them. To demonstrate the value of our approach, we present results from a 30-person lab study conducted to compare four engagement algorithms in single and multi-user scenarios. We found that combining intention to interact with a 'raise an open hand in front of you' gesture yielded the best results. The latter approach offers a 12% improvement in accuracy and a 20% reduction in time to engage over a baseline 'wave to engage' gesture currently used on the Xbox 360.

The launch of Xbox Kinect has built a very successful computer vision product and made a big impact on the gaming industry. This sheds lights onto a wide variety of potential applications related to action recognition. The accurate estimation of human poses from the depth image is universally a critical step. However, existing pose estimation systems exhibit failures when facing severe occlusion. In this paper, we propose an exemplar-based method to learn to correct the initially estimated poses. We learn an inhomogeneous systematic bias by leveraging the exemplar information within a specific human action domain. Furthermore, as an extension, we learn a conditional model by incorporation of pose tags to further increase the accuracy of pose correction. In the experiments, significant improvements on both joint-based skeleton correction and tag prediction are observed over the contemporary approaches, including what is delivered by the current Kinect system. Our experiments for the facial landmark correction also illustrate that our algorithm can improve the accuracy of other detection/estimation systems.

The launch of Xbox Kinect has built a very successful computer vision product and made a big impact to the gaming industry; this sheds lights onto a wide variety of potential applications related to action recognition. The accurate estimation of human poses from the depth image is universally a critical step. However, existing pose estimation systems exhibit failures when faced severe occlusion. In this paper, we propose an exemplar-based method to learn to correct the initially estimated poses. We learn an inhomogeneous systematic bias by leveraging the exemplar information within specific human action domain. Our algorithm is illustrated on both joint-based skeleton correction and tag prediction. In the experiments, significant improvement is observed over the contemporary approaches, including what is delivered by the current Kinect system.

When human raters are presented with a collection of shapes and asked to rank them according to their aesthetic appeal, the results often indicate that there is a statistical consensus among the raters. Yet it might be difficult to define a succinct set of rules that capture the aesthetic preferences of the raters. In this work, we explore a data-driven approach to aesthetic enhancement of such shapes. Specifically, we focus on the challenging problem of enhancing the aesthetic appeal (or the attractiveness ) of human faces in frontal photographs (portraits), while maintaining close similarity with the original. The key component in our approach is an automatic facial attractiveness engine trained on datasets of faces with accompanying facial attractiveness ratings collected from groups of human raters. Given a new face, we extract a set of distances between a variety of facial feature locations, which define a point in a high-dimensional "face space". We then search the face space for a nearby point with a higher predicted attractiveness rating. Once such a point is found, the corresponding facial distances are embedded in the plane and serve as a target to define a 2D warp field which maps the original facial features to their adjusted locations. The effectiveness of our technique was experimentally validated by independent rating experiments, which indicate that it is indeed capable of increasing the facial attractiveness of most portraits that we have experimented with.

This work presents a method for estimating human facial attractiveness, based on supervised learning techniques. Numerous facial features that describe facial geometry, color and texture, combined with an average human attractiveness score for each facial image, are used to train various predictors. Facial attractiveness ratings produced by the final predictor are found to be highly correlated with human ratings, markedly improving previous machine learning achievements. Simulated psychophysical experiments with virtually manipulated images reveal preferences in the machine's judgments which are remarkably similar to those of humans. These experiments shed new light on existing theories of facial attractiveness such as the averageness, smoothness and symmetry hypotheses. It is intriguing to find that a machine trained explicitly to capture an operational performance criteria such as attractiveness rating, implicitly captures basic human psychophysical biases characterizing the perception of facial attractiveness in general.

Harmonic colors are sets of colors that are aesthetically pleasing in terms of human visual perception. In this paper, we present a method that enhances the harmony among the colors of a given photograph or of a general image, while remaining faithful, as much as possible, to the original colors. Given a color image, our method finds the best harmonic scheme for the image colors. It then allows a graceful shifting of hue values so as to fit the harmonic scheme while considering spatial coherence among colors of neighboring pixels using an optimization technique. The results demonstrate that our method is capable of automatically enhancing the color "look-and-feel" of an ordinary image. In particular, we show the results of harmonizing the background image to accommodate the colors of a foreground image, or the foreground with respect to the background, in a cut-and-paste setting. Our color harmonization technique proves to be useful in adjusting the colors of an image composed of several parts taken from different sources.

From-region visibility culling is considered harder than from-point visibility culling, since it is inherently four-dimensional. We present a conservative occlusion culling method based on factorizing the 4D visibility problem into horizontal and vertical components. The visibility of the two components is solved asymmetrically: the horizontal component is based on a parameterization of the ray space, and the visibility of the vertical component is solved by incrementally merging umbrae. The technique is designed so that the horizontal and vertical operations can be efficiently realized together by modern graphics hardware. Similar to image-based from-point methods, we use an occlusion map to encode visibility; however, the image-space occlusion map is in the ray space rather than in the primal space. Our results show that the culling time and the size of the computed potentially visible set depend on the size of the viewcell. For moderate viewcells, conservative occlusion culling of large urban scenes takes less than a second, and the size of the potentially visible set is only about two times larger than the size of the exact visible set.