We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. By a carefully crafted design, we increased the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC14 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

/pdf/going-deeper-with-convolutions-1yobw2o2ds.pdf

Going deeper with convolutions

We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model.

/pdf/microsoft-coco-common-objects-in-context-55ihxq1zt5.pdf

Microsoft COCO: Common Objects in Context

We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Moreover, since the release of the pix2pix software associated with this paper, hundreds of twitter users have posted their own artistic experiments using our system. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without handengineering our loss functions either.

/pdf/image-to-image-translation-with-conditional-adversarial-5058t24et2.pdf

Image-to-Image Translation with Conditional Adversarial Networks

Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples. Our goal is to learn a mapping G : X → Y such that the distribution of images from G(X) is indistinguishable from the distribution Y using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping F : Y → X and introduce a cycle consistency loss to push F(G(X)) ≈ X (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach.

Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks

We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Indeed, since the release of the pix2pix software associated with this paper, a large number of internet users (many of them artists) have posted their own experiments with our system, further demonstrating its wide applicability and ease of adoption without the need for parameter tweaking. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without hand-engineering our loss functions either.

We analyze the problem of reconstructing a 2D function that approximates a set of desired gradients and a data term. The combined data and gradient terms enable operations like modifying the gradients of an image while staying close to the original image. Starting with a variational formulation, we arrive at the "screened Poisson equation" known in physics. Analysis of this equation in the Fourier domain leads to a direct, exact, and efficient solution to the problem. Further analysis reveals the structure of the spatial filters that solve the 2D screened Poisson equation and shows gradient scaling to be a well-defined sharpen filter that generalizes Laplacian sharpening, which itself can be mapped to gradient domain filtering. Results using a DCT-based screened Poisson solver are demonstrated on several applications including image blending for panoramas, image sharpening, and de-blocking of compressed images.

/pdf/fourier-analysis-of-the-2d-screened-poisson-equation-for-obz30xehv0.pdf

Fourier Analysis of the 2D Screened Poisson Equation for Gradient Domain Problems

In this paper, we explore the problem of enhancing still pictures with subtly animated motions. We limit our domain to scenes containing passive elements that respond to natural forces in some fashion. We use a semi-automatic approach, in which a human user segments the scene into a series of layers to be individually animated. Then, a "stochastic motion texture" is automatically synthesized using a spectral method, i.e., the inverse Fourier transform of a filtered noise spectrum. The motion texture is a time-varying 2D displacement map, which is applied to each layer. The resulting warped layers are then recomposited to form the animated frames. The result is a looping video texture created from a single still image, which has the advantages of being more controllable and of generally higher image quality and resolution than a video texture created from a video source. We demonstrate the technique on a variety of photographs and paintings.

Animating pictures with stochastic motion textures

We present a framework for automatically enhancing videos of a static scene using a few photographs of the same scene. For example, our system can transfer photographic qualities such as high resolution, high dynamic range and better lighting from the photographs to the video. Additionally, the user can quickly modify the video by editing only a few still images of the scene. Finally, our system allows a user to remove unwanted objects and camera shake from the video. These capabilities are enabled by two technical contributions presented in this paper. First, we make several improvements to a state-of-the-art multiview stereo algorithm in order to compute view-dependent depths using video, photographs, and structure-from-motion data. Second, we present a novel image-based rendering algorithm that can re-render the input video using the appearance of the photographs while preserving certain temporal dynamics such as specularities and dynamic scene lighting.

Using photographs to enhance videos of a static scene

We present a system for producing 3D animations using physical objects (i.e., puppets) as input. Puppeteers can load 3D models of familiar rigid objects, including toys, into our system and use them as puppets for an animation. During a performance, the puppeteer physically manipulates these puppets in front of a Kinect depth sensor. Our system uses a combination of image-feature matching and 3D shape matching to identify and track the physical puppets. It then renders the corresponding 3D models into a virtual set. Our system operates in real time so that the puppeteer can immediately see the resulting animation and make adjustments on the fly. It also provides 6D virtual camera \\rev{and lighting} controls, which the puppeteer can adjust before, during, or after a performance. Finally our system supports layered animations to help puppeteers produce animations in which several characters move at the same time. We demonstrate the accessibility of our system with a variety of animations created by puppeteers with no prior animation experience.

3D puppetry: a kinect-based interface for 3D animation

Each year, we see a growing number of 3D range scanning products on the SIGGRAPH exhibition floor. You may find yourself asking "how do these technologies work?" and "how can I make use of the shape data they produce?" In this article, I will describe a few of the more common range scanning technologies. Then, I will step through a pipeline that takes the range data into a single geometric model and will conclude with a discussion of the future of range scanning.

/pdf/from-range-scans-to-3d-models-2xhtswduty.pdf

Brian Curless

Papers

Fourier Analysis of the 2D Screened Poisson Equation for Gradient Domain Problems

Animating pictures with stochastic motion textures

Using photographs to enhance videos of a static scene

3D puppetry: a kinect-based interface for 3D animation

From range scans to 3D models