David Forsyth

Mantas Mazeika, Eric Tang, Andy Zou et al. · berkeley, cmu

In recent years, deep neural networks have demonstrated increasingly strong abilities to recognize objects and activities in videos. However, as video understanding becomes widely used in real-world applications, a key consideration is developing human-centric systems that understand not only the content of the video but also how it would affect the wellbeing and emotional state of viewers. To facilitate research in this setting, we introduce two large-scale datasets with over 60,000 videos manually annotated for emotional response and subjective wellbeing. The Video Cognitive Empathy (VCE) dataset contains annotations for distributions of fine-grained emotional responses, allowing models to gain a detailed understanding of affective states. The Video to Valence (V2V) dataset contains annotations of relative pleasantness between videos, which enables predicting a continuous spectrum of wellbeing. In experiments, we show how video models that are primarily trained to recognize actions and find contours of objects can be repurposed to understand human preferences and the emotional content of videos. Although there is room for improvement, predicting wellbeing and emotional response is on the horizon for state-of-the-art models. We hope our datasets can help foster further advances at the intersection of commonsense video understanding and human preference learning.

Yuzheng Hu, Fan Wu, Qinbin Li et al.

As the prevalence of data analysis grows, safeguarding data privacy has become a paramount concern. Consequently, there has been an upsurge in the development of mechanisms aimed at privacy-preserving data analyses. However, these approaches are task-specific; designing algorithms for new tasks is a cumbersome process. As an alternative, one can create synthetic data that is (ideally) devoid of private information. This paper focuses on privacy-preserving data synthesis (PPDS) by providing a comprehensive overview, analysis, and discussion of the field. Specifically, we put forth a master recipe that unifies two prominent strands of research in PPDS: statistical methods and deep learning (DL)-based methods. Under the master recipe, we further dissect the statistical methods into choices of modeling and representation, and investigate the DL-based methods by different generative modeling principles. To consolidate our findings, we provide comprehensive reference tables, distill key takeaways, and identify open problems in the existing literature. In doing so, we aim to answer the following questions: What are the design principles behind different PPDS methods? How can we categorize these methods, and what are the advantages and disadvantages associated with each category? Can we provide guidelines for method selection in different real-world scenarios? We proceed to benchmark several prominent DL-based methods on the task of private image synthesis and conclude that DP-MERF is an all-purpose approach. Finally, upon systematizing the work over the past decade, we identify future directions and call for actions from researchers.

Mantas Mazeika, Bo Li, David Forsyth

Model stealing attacks present a dilemma for public machine learning APIs. To protect financial investments, companies may be forced to withhold important information about their models that could facilitate theft, including uncertainty estimates and prediction explanations. This compromise is harmful not only to users but also to external transparency. Model stealing defenses seek to resolve this dilemma by making models harder to steal while preserving utility for benign users. However, existing defenses have poor performance in practice, either requiring enormous computational overheads or severe utility trade-offs. To meet these challenges, we present a new approach to model stealing defenses called gradient redirection. At the core of our approach is a provably optimal, efficient algorithm for steering an adversary's training updates in a targeted manner. Combined with improvements to surrogate networks and a novel coordinated defense strategy, our gradient redirection defense, called GRAD${}^2$, achieves small utility trade-offs and low computational overhead, outperforming the best prior defenses. Moreover, we demonstrate how gradient redirection enables reprogramming the adversary with arbitrary behavior, which we hope will foster work on new avenues of defense.

Yuanyi Zhong, Anand Bhattad, Yu-Xiong Wang et al.

Dense depth and surface normal predictors should possess the equivariant property to cropping-and-resizing -- cropping the input image should result in cropping the same output image. However, we find that state-of-the-art depth and normal predictors, despite having strong performances, surprisingly do not respect equivariance. The problem exists even when crop-and-resize data augmentation is employed during training. To remedy this, we propose an equivariant regularization technique, consisting of an averaging procedure and a self-consistency loss, to explicitly promote cropping-and-resizing equivariance in depth and normal networks. Our approach can be applied to both CNN and Transformer architectures, does not incur extra cost during testing, and notably improves the supervised and semi-supervised learning performance of dense predictors on Taskonomy tasks. Finally, finetuning with our loss on unlabeled images improves not only equivariance but also accuracy of state-of-the-art depth and normal predictors when evaluated on NYU-v2. GitHub link: https://github.com/mikuhatsune/equivariance

Yuan Li, Zhi-Hao Lin, David Forsyth et al.

Physical simulations produce excellent predictions of weather effects. Neural radiance fields produce SOTA scene models. We describe a novel NeRF-editing procedure that can fuse physical simulations with NeRF models of scenes, producing realistic movies of physical phenomena in those scenes. Our application -- Climate NeRF -- allows people to visualize what climate change outcomes will do to them. ClimateNeRF allows us to render realistic weather effects, including smog, snow, and flood. Results can be controlled with physically meaningful variables like water level. Qualitative and quantitative studies show that our simulated results are significantly more realistic than those from SOTA 2D image editing and SOTA 3D NeRF stylization.

Chih-Hao Lin, Bohan Liu, Yi-Ting Chen et al.

We present UrbanIR (Urban Scene Inverse Rendering), a new inverse graphics model that enables realistic, free-viewpoint renderings of scenes under various lighting conditions with a single video. It accurately infers shape, albedo, visibility, and sun and sky illumination from wide-baseline videos, such as those from car-mounted cameras, differing from NeRF's dense view settings. In this context, standard methods often yield subpar geometry and material estimates, such as inaccurate roof representations and numerous 'floaters'. UrbanIR addresses these issues with novel losses that reduce errors in inverse graphics inference and rendering artifacts. Its techniques allow for precise shadow volume estimation in the original scene. The model's outputs support controllable editing, enabling photorealistic free-viewpoint renderings of night simulations, relit scenes, and inserted objects, marking a significant improvement over existing state-of-the-art methods.

Ayush Sarkar, Hanlin Mai, Amitabh Mahapatra et al.

Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors, for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images.

Mantas Mazeika, Long Phan, Xuwang Yin et al. · berkeley, cmu

Automated red teaming holds substantial promise for uncovering and mitigating the risks associated with the malicious use of large language models (LLMs), yet the field lacks a standardized evaluation framework to rigorously assess new methods. To address this issue, we introduce HarmBench, a standardized evaluation framework for automated red teaming. We identify several desirable properties previously unaccounted for in red teaming evaluations and systematically design HarmBench to meet these criteria. Using HarmBench, we conduct a large-scale comparison of 18 red teaming methods and 33 target LLMs and defenses, yielding novel insights. We also introduce a highly efficient adversarial training method that greatly enhances LLM robustness across a wide range of attacks, demonstrating how HarmBench enables codevelopment of attacks and defenses. We open source HarmBench at https://github.com/centerforaisafety/HarmBench.

Vaibhav Vavilala, Seemandhar Jain, Rahul Vasanth et al.

We present Blocks2World, a novel method for 3D scene rendering and editing that leverages a two-step process: convex decomposition of images and conditioned synthesis. Our technique begins by extracting 3D parallelepipeds from various objects in a given scene using convex decomposition, thus obtaining a primitive representation of the scene. These primitives are then utilized to generate paired data through simple ray-traced depth maps. The next stage involves training a conditioned model that learns to generate images from the 2D-rendered convex primitives. This step establishes a direct mapping between the 3D model and its 2D representation, effectively learning the transition from a 3D model to an image. Once the model is fully trained, it offers remarkable control over the synthesis of novel and edited scenes. This is achieved by manipulating the primitives at test time, including translating or adding them, thereby enabling a highly customizable scene rendering process. Our method provides a fresh perspective on 3D scene rendering and editing, offering control and flexibility. It opens up new avenues for research and applications in the field, including authoring and data augmentation.

Vaibhav Vavilala, David Forsyth

We describe a method to parse a complex, cluttered indoor scene into primitives which offer a parsimonious abstraction of scene structure. Our primitives are simple convexes. Our method uses a learned regression procedure to parse a scene into a fixed number of convexes from RGBD input, and can optionally accept segmentations to improve the decomposition. The result is then polished with a descent method which adjusts the convexes to produce a very good fit, and greedily removes superfluous primitives. Because the entire scene is parsed, we can evaluate using traditional depth, normal, and segmentation error metrics. Our evaluation procedure demonstrates that the error from our primitive representation is comparable to that of predicting depth from a single image.

Sara Aghajanzadeh, David Forsyth

Image frames obtained in darkness are special. Just multiplying by a constant doesn't restore the image. Shot noise, quantization effects and camera non-linearities mean that colors and relative light levels are estimated poorly. Current methods learn a mapping using real dark-bright image pairs. These are very hard to capture. A recent paper has shown that simulated data pairs produce real improvements in restoration, likely because huge volumes of simulated data are easy to obtain. In this paper, we show that respecting equivariance -- the color of a restored pixel should be the same, however the image is cropped -- produces real improvements over the state of the art for restoration. We show that a scale selection mechanism can be used to improve reconstructions. Finally, we show that our approach produces improvements on video restoration as well. Our methods are evaluated both quantitatively and qualitatively.

Sara Aghajanzadeh, David Forsyth

In this paper, we study the problem of making brighter images from dark images found in the wild. The images are dark because they are taken in dim environments. They suffer from color shifts caused by quantization and from sensor noise. We don't know the true camera reponse function for such images and they are not RAW. We use a supervised learning method, relying on a straightforward simulation of an imaging pipeline to generate usable dataset for training and testing. On a number of standard datasets, our approach outperforms the state of the art quantitatively. Qualitative comparisons suggest strong improvements in reconstruction accuracy.

Kedan Li, Jeffrey Zhang, Shao-Yu Chang et al.

An outfit visualization method generates an image of a person wearing real garments from images of those garments. Current methods can produce images that look realistic and preserve garment identity, captured in details such as collar, cuffs, texture, hem, and sleeve length. However, no current method can both control how the garment is worn -- including tuck or untuck, opened or closed, high or low on the waist, etc.. -- and generate realistic images that accurately preserve the properties of the original garment. We describe an outfit visualization method that controls drape while preserving garment identity. Our system allows instance independent editing of garment drape, which means a user can construct an edit (e.g. tucking a shirt in a specific way) that can be applied to all shirts in a garment collection. Garment detail is preserved by relying on a warping procedure to place the garment on the body and a generator then supplies fine shading detail. To achieve instance independent control, we use control points with garment category-level semantics to guide the warp. The method produces state-of-the-art quality images, while allowing creative ways to style garments, including allowing tops to be tucked or untucked; jackets to be worn open or closed; skirts to be worn higher or lower on the waist; and so on. The method allows interactive control to correct errors in individual renderings too. Because the edits are instance independent, they can be applied to large pools of garments automatically and can be conditioned on garment metadata (e.g. all cropped jackets are worn closed or all bomber jackets are worn closed).

Saba Ghaffari, Ehsan Saleh, David Forsyth et al.

Learning accurate classifiers for novel categories from very few examples, known as few-shot image classification, is a challenging task in statistical machine learning and computer vision. The performance in few-shot classification suffers from the bias in the estimation of classifier parameters; however, an effective underlying bias reduction technique that could alleviate this issue in training few-shot classifiers has been overlooked. In this work, we demonstrate the effectiveness of Firth bias reduction in few-shot classification. Theoretically, Firth bias reduction removes the $O(N^{-1})$ first order term from the small-sample bias of the Maximum Likelihood Estimator. Here we show that the general Firth bias reduction technique simplifies to encouraging uniform class assignment probabilities for multinomial logistic classification, and almost has the same effect in cosine classifiers. We derive an easy-to-implement optimization objective for Firth penalized multinomial logistic and cosine classifiers, which is equivalent to penalizing the cross-entropy loss with a KL-divergence between the uniform label distribution and the predictions. Then, we empirically evaluate that it is consistently effective across the board for few-shot image classification, regardless of (1) the feature representations from different backbones, (2) the number of samples per class, and (3) the number of classes. Finally, we show the robustness of Firth bias reduction, in the case of imbalanced data distribution. Our implementation is available at https://github.com/ehsansaleh/firth_bias_reduction

Bryan A. Plummer, Mariya I. Vasileva, Vitali Petsiuk et al.

Explaining a deep learning model can help users understand its behavior and allow researchers to discern its shortcomings. Recent work has primarily focused on explaining models for tasks like image classification or visual question answering. In this paper, we introduce Salient Attributes for Network Explanation (SANE) to explain image similarity models, where a model's output is a score measuring the similarity of two inputs rather than a classification score. In this task, an explanation depends on both of the input images, so standard methods do not apply. Our SANE explanations pairs a saliency map identifying important image regions with an attribute that best explains the match. We find that our explanations provide additional information not typically captured by saliency maps alone, and can also improve performance on the classic task of attribute recognition. Our approach's ability to generalize is demonstrated on two datasets from diverse domains, Polyvore Outfits and Animals with Attributes 2. Code available at: https://github.com/VisionLearningGroup/SANE

Aditya Prakash, Benjamin Lundell, Dmitry Andreychuk et al.

We tackle the novel problem of predicting 3D hand motion and contact maps (or Interaction Trajectories) given a single RGB view, action text, and a 3D contact point on the object as input. Our approach consists of (1) Interaction Codebook: a VQVAE model to learn a latent codebook of hand poses and contact points, effectively tokenizing interaction trajectories, (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs by using an indexer module to retrieve a latent affordance from the learned codebook. To train our model, we develop a data engine that extracts 3D hand poses and contact trajectories from the diverse HoloAssist dataset. We evaluate our model on a benchmark that is 2.5-10X larger than existing works, in terms of diversity of objects and interactions observed, and test for generalization of the model across object categories, action categories, tasks, and scenes. Experimental results show the effectiveness of our approach over transformer & diffusion baselines across all settings.

Jeffrey Zhang, Shao-Yu Chang, Kedan Li et al.

Retail photography imposes specific requirements on images. For instance, images may need uniform background colors, consistent model poses, centered products, and consistent lighting. Minor deviations from these standards impact a site's aesthetic appeal, making the images unsuitable for use. We show that Stable Diffusion methods, as currently applied, do not respect these requirements. The usual practice of training the denoiser with a very noisy image and starting inference with a sample of pure noise leads to inconsistent generated images during inference. This inconsistency occurs because it is easy to tell the difference between samples of the training and inference distributions. As a result, a network trained with centered retail product images with uniform backgrounds generates images with erratic backgrounds. The problem is easily fixed by initializing inference with samples from an approximation of noisy images. However, in using such an approximation, the joint distribution of text and noisy image at inference time still slightly differs from that at training time. This discrepancy is corrected by training the network with samples from the approximate noisy image distribution. Extensive experiments on real application data show significant qualitative and quantitative improvements in performance from adopting these procedures. Finally, our procedure can interact well with other control-based methods to further enhance the controllability of diffusion-based methods.

Vaibhav Vavilala, Rahul Vasanth, David Forsyth

Physically-based renderings contain Monte-Carlo noise, with variance that increases as the number of rays per pixel decreases. This noise, while zero-mean for good modern renderers, can have heavy tails (most notably, for scenes containing specular or refractive objects). Learned methods for restoring low fidelity renders are highly developed, because suppressing render noise means one can save compute and use fast renders with few rays per pixel. We demonstrate that a diffusion model can denoise low fidelity renders successfully. Furthermore, our method can be conditioned on a variety of natural render information, and this conditioning helps performance. Quantitative experiments show that our method is competitive with SOTA across a range of sampling rates. Qualitative examination of the reconstructions suggests that the image prior applied by a diffusion method strongly favors reconstructions that are like real images -- so have straight shadow boundaries, curved specularities and no fireflies.

Joshua Cho, Sara Aghajanzadeh, Zhen Zhu et al.

In this paper, we present a simple yet highly effective "free lunch" solution for low-light image enhancement (LLIE), which aims to restore low-light images as if acquired in well-illuminated environments. Our method necessitates no optimization, training, fine-tuning, text conditioning, or hyperparameter adjustments, yet it consistently reconstructs low-light images with superior fidelity. Specifically, we leverage a pre-trained text-to-image diffusion prior, learned from training on a large collection of natural images, and the features present in the model itself to guide the inference, in contrast to existing methods that depend on customized constraints. Comprehensive quantitative evaluations demonstrate that our approach outperforms SOTA methods on established datasets, while qualitative analyses indicate enhanced color accuracy and the rectification of subtle chromatic deviations. Furthermore, additional experiments reveal that our method, without any modifications, achieves SOTA-comparable performance in the auto white balance (AWB) task.

Jeffrey Zhang, Kedan Li, Shao-Yu Chang et al.

Virtual Try-on (VTON) involves generating images of a person wearing selected garments. Diffusion-based methods, in particular, can create high-quality images, but they struggle to maintain the identities of the input garments. We identified this problem stems from the specifics in the training formulation for diffusion. To address this, we propose a unique training scheme that limits the scope in which diffusion is trained. We use a control image that perfectly aligns with the target image during training. In turn, this accurately preserves garment details during inference. We demonstrate our method not only effectively conserves garment details but also allows for layering, styling, and shoe try-on. Our method runs multi-garment try-on in a single inference cycle and can support high-quality zoomed-in generations without training in higher resolutions. Finally, we show our method surpasses prior methods in accuracy and quality.

Aditya Prakash, David Forsyth, Saurabh Gupta

We tackle the problem of forecasting bimanual 3D hand motion & articulation from a single image in everyday settings. To address the lack of 3D hand annotations in diverse settings, we design an annotation pipeline consisting of a diffusion model to lift 2D hand keypoint sequences to 4D hand motion. For the forecasting model, we adopt a diffusion loss to account for the multimodality in hand motion distribution. Extensive experiments across 6 datasets show the benefits of training on diverse data with imputed labels (14% improvement) and effectiveness of our lifting (42% better) & forecasting (16.4% gain) models, over the best baselines, especially in zero-shot generalization to everyday images.

Xiaoxue Chen, Bhargav Chandaka, Chih-Hao Lin et al.

We present InvRGB+L, a novel inverse rendering model that reconstructs large, relightable, and dynamic scenes from a single RGB+LiDAR sequence. Conventional inverse graphics methods rely primarily on RGB observations and use LiDAR mainly for geometric information, often resulting in suboptimal material estimates due to visible light interference. We find that LiDAR's intensity values-captured with active illumination in a different spectral range-offer complementary cues for robust material estimation under variable lighting. Inspired by this, InvRGB+L leverages LiDAR intensity cues to overcome challenges inherent in RGB-centric inverse graphics through two key innovations: (1) a novel physics-based LiDAR shading model and (2) RGB-LiDAR material consistency losses. The model produces novel-view RGB and LiDAR renderings of urban and indoor scenes and supports relighting, night simulations, and dynamic object insertions, achieving results that surpass current state-of-the-art methods in both scene-level urban inverse rendering and LiDAR simulation.

Min Jin Chong, David Forsyth

A style mapper applies some fixed style to its input images (so, for example, taking faces to cartoons). This paper describes a simple procedure -- JoJoGAN -- to learn a style mapper from a single example of the style. JoJoGAN uses a GAN inversion procedure and StyleGAN's style-mixing property to produce a substantial paired dataset from a single example style. The paired dataset is then used to fine-tune a StyleGAN. An image can then be style mapped by GAN-inversion followed by the fine-tuned StyleGAN. JoJoGAN needs just one reference and as little as 30 seconds of training time. JoJoGAN can use extreme style references (say, animal faces) successfully. Furthermore, one can control what aspects of the style are used and how much of the style is applied. Qualitative and quantitative evaluation show that JoJoGAN produces high quality high resolution images that vastly outperform the current state-of-the-art.

Liwen Wu, Jae Yong Lee, Anand Bhattad et al.

DIVeR builds on the key ideas of NeRF and its variants -- density models and volume rendering -- to learn 3D object models that can be rendered realistically from small numbers of images. In contrast to all previous NeRF methods, DIVeR uses deterministic rather than stochastic estimates of the volume rendering integral. DIVeR's representation is a voxel based field of features. To compute the volume rendering integral, a ray is broken into intervals, one per voxel; components of the volume rendering integral are estimated from the features for each interval using an MLP, and the components are aggregated. As a result, DIVeR can render thin translucent structures that are missed by other integrators. Furthermore, DIVeR's representation has semantics that is relatively exposed compared to other such methods -- moving feature vectors around in the voxel space results in natural edits. Extensive qualitative and quantitative comparisons to current state-of-the-art methods show that DIVeR produces models that (1) render at or above state-of-the-art quality, (2) are very small without being baked, (3) render very fast without being baked, and (4) can be edited in natural ways.

Min Jin Chong, Hsin-Ying Lee, David Forsyth

Recently, StyleGAN has enabled various image manipulation and editing tasks thanks to the high-quality generation and the disentangled latent space. However, additional architectures or task-specific training paradigms are usually required for different tasks. In this work, we take a deeper look at the spatial properties of StyleGAN. We show that with a pretrained StyleGAN along with some operations, without any additional architecture, we can perform comparably to the state-of-the-art methods on various tasks, including image blending, panorama generation, generation from a single image, controllable and local multimodal image to image translation, and attributes transfer. The proposed method is simple, effective, efficient, and applicable to any existing pretrained StyleGAN model.

Vaibhav Vavilala, David Forsyth

The state-of-the-art StyleGAN2 network supports powerful methods to create and edit art, including generating random images, finding images "like" some query, and modifying content or style. Further, recent advancements enable training with small datasets. We apply these methods to synthesize card art, by training on a novel Yu-Gi-Oh dataset. While noise inputs to StyleGAN2 are essential for good synthesis, we find that coarse-scale noise interferes with latent variables on this dataset because both control long-scale image effects. We observe over-aggressive variation in art with changes in noise and weak content control via latent variable edits. Here, we demonstrate that training a modified StyleGAN2, where coarse-scale noise is suppressed, removes these unwanted effects. We obtain a superior FID; changes in noise result in local exploration of style; and identity control is markedly improved. These results and analysis lead towards a GAN-assisted art synthesis tool for digital artists of all skill levels, which can be used in film, games, or any creative industry for artistic ideation.

Dominic Roberts, Ara Danielyan, Hang Chu et al.

Generative models for 3D shapes represented by hierarchies of parts can generate realistic and diverse sets of outputs. However, existing models suffer from the key practical limitation of modelling shapes holistically and thus cannot perform conditional sampling, i.e. they are not able to generate variants on individual parts of generated shapes without modifying the rest of the shape. This is limiting for applications such as 3D CAD design that involve adjusting created shapes at multiple levels of detail. To address this, we introduce LSD-StructureNet, an augmentation to the StructureNet architecture that enables re-generation of parts situated at arbitrary positions in the hierarchies of its outputs. We achieve this by learning individual, probabilistic conditional decoders for each hierarchy depth. We evaluate LSD-StructureNet on the PartNet dataset, the largest dataset of 3D shapes represented by hierarchies of parts. Our results show that contrarily to existing methods, LSD-StructureNet can perform conditional sampling without impacting inference speed or the realism and diversity of its outputs.

Min Jin Chong, Wen-Sheng Chu, Abhishek Kumar et al.

We present Retrieve in Style (RIS), an unsupervised framework for facial feature transfer and retrieval on real images. Recent work shows capabilities of transferring local facial features by capitalizing on the disentanglement property of the StyleGAN latent space. RIS improves existing art on the following: 1) Introducing more effective feature disentanglement to allow for challenging transfers (ie, hair, pose) that were not shown possible in SoTA methods. 2) Eliminating the need for per-image hyperparameter tuning, and for computing a catalog over a large batch of images. 3) Enabling fine-grained face retrieval using disentangled facial features (eg, eyes). To our best knowledge, this is the first work to retrieve face images at this fine level. 4) Demonstrating robust, natural editing on real images. Our qualitative and quantitative analyses show RIS achieves both high-fidelity feature transfers and accurate fine-grained retrievals on real images. We also discuss the responsible applications of RIS.

Min Jin Chong, David Forsyth

We show how to learn a map that takes a content code, derived from a face image, and a randomly chosen style code to an anime image. We derive an adversarial loss from our simple and effective definitions of style and content. This adversarial loss guarantees the map is diverse -- a very wide range of anime can be produced from a single content code. Under plausible assumptions, the map is not just diverse, but also correctly represents the probability of an anime, conditioned on an input face. In contrast, current multimodal generation procedures cannot capture the complex styles that appear in anime. Extensive quantitative experiments support the idea the map is correct. Extensive qualitative results show that the method can generate a much more diverse range of styles than SOTA comparisons. Finally, we show that our formalization of content and style allows us to perform video to video translation without ever training on videos.

D. A. Forsyth, Jason J. Rock

Intrinsic image decomposition is the classical task of mapping image to albedo. The WHDR dataset allows methods to be evaluated by comparing predictions to human judgements ("lighter", "same as", "darker"). The best modern intrinsic image methods learn a map from image to albedo using rendered models and human judgements. This is convenient for practical methods, but cannot explain how a visual agent without geometric, surface and illumination models and a renderer could learn to recover intrinsic images. This paper describes a method that learns intrinsic image decomposition without seeing WHDR annotations, rendered data, or ground truth data. The method relies on paradigms - fake albedos and fake shading fields - together with a novel smoothing procedure that ensures good behavior at short scales on real images. Long scale error is controlled by averaging. Our method achieves WHDR scores competitive with those of strong recent methods allowed to see training WHDR annotations, rendered data, and ground truth data. Because our method is unsupervised, we can compute estimates of the test/train variance of WHDR scores; these are quite large, and it is unsafe to rely small differences in reported WHDR.

Weilin Zhang, Yu-Xiong Wang, David A. Forsyth

Learning to detect an object in an image from very few training examples - few-shot object detection - is challenging, because the classifier that sees proposal boxes has very little training data. A particularly challenging training regime occurs when there are one or two training examples. In this case, if the region proposal network (RPN) misses even one high intersection-over-union (IOU) training box, the classifier's model of how object appearance varies can be severely impacted. We use multiple distinct yet cooperating RPN's. Our RPN's are trained to be different, but not too different; doing so yields significant performance improvements over state of the art for COCO and PASCAL VOC in the very few-shot setting. This effect appears to be independent of the choice of classifier or dataset.

Anand Bhattad, David A. Forsyth

We show how to insert an object from one image to another and get realistic results in the hard case, where the shading of the inserted object clashes with the shading of the scene. Rendering objects using an illumination model of the scene doesn't work, because doing so requires a geometric and material model of the object, which is hard to recover from a single image. In this paper, we introduce a method that corrects shading inconsistencies of the inserted object without requiring a geometric and physical model or an environment map. Our method uses a deep image prior (DIP), trained to produce reshaded renderings of inserted objects via consistent image decomposition inferential losses. The resulting image from DIP aims to have (a) an albedo similar to the cut-and-paste albedo, (b) a similar shading field to that of the target scene, and (c) a shading that is consistent with the cut-and-paste surface normals. The result is a simple procedure that produces convincing shading of the inserted object. We show the efficacy of our method both qualitatively and quantitatively for several objects with complex surface properties and also on a dataset of spherical lampshades for quantitative evaluation. Our method significantly outperforms an Image Harmonization (IH) baseline for all these objects. They also outperform the cut-and-paste and IH baselines in a user study with over 100 users.

Kedan Li, Min Jin Chong, Jingen Liu et al.

A virtual try-on method takes a product image and an image of a model and produces an image of the model wearing the product. Most methods essentially compute warps from the product image to the model image and combine using image generation methods. However, obtaining a realistic image is challenging because the kinematics of garments is complex and because outline, texture, and shading cues in the image reveal errors to human viewers. The garment must have appropriate drapes; texture must be warped to be consistent with the shape of a draped garment; small details (buttons, collars, lapels, pockets, etc.) must be placed appropriately on the garment, and so on. Evaluation is particularly difficult and is usually qualitative. This paper uses quantitative evaluation on a challenging, novel dataset to demonstrate that (a) for any warping method, one can choose target models automatically to improve results, and (b) learning multiple coordinated specialized warpers offers further improvements on results. Target models are chosen by a learned embedding procedure that predicts a representation of the products the model is wearing. This prediction is used to match products to models. Specialized warpers are trained by a method that encourages a second warper to perform well in locations where the first works poorly. The warps are then combined using a U-Net. Qualitative evaluation confirms that these improvements are wholesale over outline, texture shading, and garment details.

Shruti Bhargava, David Forsyth

The task of image captioning implicitly involves gender identification. However, due to the gender bias in data, gender identification by an image captioning model suffers. Also, the gender-activity bias, owing to the word-by-word prediction, influences other words in the caption prediction, resulting in the well-known problem of label bias. In this work, we investigate gender bias in the COCO captioning dataset and show that it engenders not only from the statistical distribution of genders with contexts but also from the flawed annotation by the human annotators. We look at the issues created by this bias in the trained models. We propose a technique to get rid of the bias by splitting the task into 2 subtasks: gender-neutral image captioning and gender classification. By this decoupling, the gender-context influence can be eradicated. We train the gender-neutral image captioning model, which gives comparable results to a gendered model even when evaluating against a dataset that possesses a similar bias as the training data. Interestingly, the predictions by this model on images with no humans, are also visibly different from the one trained on gendered captions. We train gender classifiers using the available bounding box and mask-based annotations for the person in the image. This allows us to get rid of the context and focus on the person to predict the gender. By substituting the genders into the gender-neutral captions, we get the final gendered predictions. Our predictions achieve similar performance to a model trained with gender, and at the same time are devoid of gender bias. Finally, our main result is that on an anti-stereotypical dataset, our model outperforms a popular image captioning model which is trained with gender.

Min Jin Chong, David Forsyth

This paper shows that two commonly used evaluation metrics for generative models, the Fréchet Inception Distance (FID) and the Inception Score (IS), are biased -- the expected value of the score computed for a finite sample set is not the true value of the score. Worse, the paper shows that the bias term depends on the particular model being evaluated, so model A may get a better score than model B simply because model A's bias term is smaller. This effect cannot be fixed by evaluating at a fixed number of samples. This means all comparisons using FID or IS as currently computed are unreliable. We then show how to extrapolate the score to obtain an effectively bias-free estimate of scores computed with an infinite number of samples, which we term $\overline{\textrm{FID}}_\infty$ and $\overline{\textrm{IS}}_\infty$. In turn, this effectively bias-free estimate requires good estimates of scores with a finite number of samples. We show that using Quasi-Monte Carlo integration notably improves estimates of FID and IS for finite sample sets. Our extrapolated scores are simple, drop-in replacements for the finite sample scores. Additionally, we show that using low discrepancy sequence in GAN training offers small improvements in the resulting generator.

Mao-Chuang Yeh, Shuai Tang, Anand Bhattad et al.

Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. We seek to understand how to improve style transfer. To do so requires quantitative evaluation procedures, but the current evaluation is qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. Our procedure relies on two statistics: the Effectiveness (E) statistic measures the extent that a given style has been transferred to the target, and the Coherence (C) statistic measures the extent to which the original image's content is preserved. Our statistics are calibrated to human preference: targets with larger values of E (resp C) will reliably be preferred by human subjects in comparisons of style (resp. content). We use these statistics to investigate the relative performance of a number of Neural Style Transfer(NST) methods, revealing several intriguing properties. Admissible methods lie on a Pareto frontier (i.e. improving E reduces C or vice versa). Three methods are admissible: Universal style transfer produces very good C but weak E; modifying the optimization used for Gatys' loss produces a method with strong E and strong C; and a modified cross-layer method has slightly better E at strong cost in C. While the histogram loss improves the E statistics of Gatys' method, it does not make the method admissible. Surprisingly, style weights have relatively little effect in improving EC scores, and most variability in the transfer is explained by the style itself (meaning experimenters can be misguided by selecting styles).

Theerasit Issaranon, Chuhang Zou, David Forsyth

We describe a method that predicts, from a single RGB image, a depth map that describes the scene when a masked object is removed - we call this "counterfactual depth" that models hidden scene geometry together with the observations. Our method works for the same reason that scene completion works: the spatial structure of objects is simple. But we offer a much higher resolution representation of space than current scene completion methods, as we operate at pixel-level precision and do not rely on a voxel representation. Furthermore, we do not require RGBD inputs. Our method uses a standard encoder-decoder architecture, and with a decoder modified to accept an object mask. We describe a small evaluation dataset that we have collected, which allows inference about what factors affect reconstruction most strongly. Using this dataset, we show that our depth predictions for masked objects are better than other baselines.

Anand Bhattad, Min Jin Chong, Kaizhao Liang et al.

Machine learning models, especially deep neural networks (DNNs), have been shown to be vulnerable against adversarial examples which are carefully crafted samples with a small magnitude of the perturbation. Such adversarial perturbations are usually restricted by bounding their $\mathcal{L}_p$ norm such that they are imperceptible, and thus many current defenses can exploit this property to reduce their adversarial impact. In this paper, we instead introduce "unrestricted" perturbations that manipulate semantically meaningful image-based visual descriptors - color and texture - in order to generate effective and photorealistic adversarial examples. We show that these semantically aware perturbations are effective against JPEG compression, feature squeezing and adversarially trained model. We also show that the proposed methods can effectively be applied to both image classification and image captioning tasks on complex datasets such as ImageNet and MSCOCO. In addition, we conduct comprehensive user studies to show that our generated semantic adversarial examples are photorealistic to humans despite large magnitude perturbations when compared to other attacks.

Ishan Deshpande, Yuan-Ting Hu, Ruoyu Sun et al.

Generative adversarial nets (GANs) and variational auto-encoders have significantly improved our distribution modeling capabilities, showing promise for dataset augmentation, image-to-image translation and feature learning. However, to model high-dimensional distributions, sequential training and stacked architectures are common, increasing the number of tunable hyper-parameters as well as the training time. Nonetheless, the sample complexity of the distance metrics remains one of the factors affecting GAN training. We first show that the recently proposed sliced Wasserstein distance has compelling sample complexity properties when compared to the Wasserstein distance. To further improve the sliced Wasserstein distance we then analyze its `projection complexity' and develop the max-sliced Wasserstein distance which enjoys compelling sample complexity while reducing projection complexity, albeit necessitating a max estimation. We finally illustrate that the proposed distance trains GANs on high-dimensional images up to a resolution of 256x256 easily.

Safa Messaoud, David Forsyth, Alexander G. Schwing

Colorizing a given gray-level image is an important task in the media and advertising industry. Due to the ambiguity inherent to colorization (many shades are often plausible), recent approaches started to explicitly model diversity. However, one of the most obvious artifacts, structural inconsistency, is rarely considered by existing methods which predict chrominance independently for every pixel. To address this issue, we develop a conditional random field based variational auto-encoder formulation which is able to achieve diversity while taking into account structural consistency. Moreover, we introduce a controllability mecha- nism that can incorporate external constraints from diverse sources in- cluding a user interface. Compared to existing baselines, we demonstrate that our method obtains more diverse and globally consistent coloriza- tions on the LFW, LSUN-Church and ILSVRC-2015 datasets.

Aditya Deshpande, Jyoti Aneja, Liwei Wang et al.

Image captioning is an ambiguous problem, with many suitable captions for an image. To address ambiguity, beam search is the de facto method for sampling multiple captions. However, beam search is computationally expensive and known to produce generic captions. To address this concern, some variational auto-encoder (VAE) and generative adversarial net (GAN) based methods have been proposed. Though diverse, GAN and VAE are less accurate. In this paper, we first predict a meaningful summary of the image, then generate the caption based on that summary. We use part-of-speech as summaries, since our summary should drive caption generation. We achieve the trifecta: (1) High accuracy for the diverse captions as evaluated by standard captioning metrics and user studies; (2) Faster computation of diverse captions compared to beam search and diverse beam search; and (3) High diversity as evaluated by counting novel sentences, distinct n-grams and mutual overlap (i.e., mBleu-4) scores.

Mao-Chuang Yeh, Shuai Tang, Anand Bhattad et al.

Style transfer methods produce a transferred image which is a rendering of a content image in the manner of a style image. There is a rich literature of variant methods. However, evaluation procedures are qualitative, mostly involving user studies. We describe a novel quantitative evaluation procedure. One plots effectiveness (a measure of the extent to which the style was transferred) against coherence (a measure of the extent to which the transferred image decomposes into objects in the same way that the content image does) to obtain an EC plot. We construct EC plots comparing a number of recent style transfer methods. Most methods control within-layer gram matrices, but we also investigate a method that controls cross-layer gram matrices. These EC plots reveal a number of intriguing properties of recent style transfer methods. The style used has a strong effect on the outcome, for all methods. Using large style weights does not necessarily improve effectiveness, and can produce worse results. Cross-layer gram matrices easily beat all other methods, but some styles remain difficult for all methods. Ensemble methods show real promise. It is likely that, for current methods, each style requires a different choice of weights to obtain the best results, so that automated weight setting methods are desirable. Finally, we show evidence comparing our EC evaluations to human evaluations.

Mariya I. Vasileva, Bryan A. Plummer, Krishna Dusad et al.

Outfits in online fashion data are composed of items of many different types (e.g. top, bottom, shoes) that share some stylistic relationship with one another. A representation for building outfits requires a method that can learn both notions of similarity (for example, when two tops are interchangeable) and compatibility (items of possibly different type that can go together in an outfit). This paper presents an approach to learning an image embedding that respects item type, and jointly learns notions of item similarity and compatibility in an end-to-end model. To evaluate the learned representation, we crawled 68,306 outfits created by users on the Polyvore website. Our approach obtains 3-5% improvement over the state-of-the-art on outfit compatibility prediction and fill-in-the-blank tasks using our dataset, as well as an established smaller dataset, while supporting a variety of useful queries.

Anand Bhattad, Jason Rock, David Forsyth

Detecting anomalous faces has important applications. For example, a system might tell when a train driver is incapacitated by a medical event, and assist in adopting a safe recovery strategy. These applications are demanding, because they require accurate detection of rare anomalies that may be seen only at runtime. Such a setting causes supervised methods to perform poorly. We describe a method for detecting an anomalous face image that meets these requirements. We construct a feature vector that reliably has large entries for anomalous images, then use various simple unsupervised methods to score the image based on the feature. Obvious constructions (autoencoder codes; autoencoder residuals) are defeated by a 'peeking' behavior in autoencoders. Our feature construction removes rectangular patches from the image, predicts the likely content of the patch conditioned on the rest of the image using a specially trained autoencoder, then compares the result to the image. High scores suggest that the patch was difficult for an autoencoder to predict, and so is likely anomalous. We demonstrate that our method can identify real anomalous face images in pools of typical images, taken from celeb-A, that is much larger than usual in state-of-the-art experiments. A control experiment based on our method with another set of normal celebrity images - a 'typical set', but nonceleb-A are not identified as anomalous; confirms this is not due to special properties of celeb-A.

Jiajun Lu, Hussein Sibai, Evan Fabry et al.

An adversarial example is an example that has been adjusted to produce the wrong label when presented to a system at test time. If adversarial examples existed that could fool a detector, they could be used to (for example) wreak havoc on roads populated with smart vehicles. Recently, we described our difficulties creating physical adversarial stop signs that fool a detector. More recently, Evtimov et al. produced a physical adversarial stop sign that fools a proxy model of a detector. In this paper, we show that these physical adversarial stop signs do not fool two standard detectors (YOLO and Faster RCNN) in standard configuration. Evtimov et al.'s construction relies on a crop of the image to the stop sign; this crop is then resized and presented to a classifier. We argue that the cropping and resizing procedure largely eliminates the effects of rescaling and of view angle. Whether an adversarial attack is robust under rescaling and change of view direction remains moot. We argue that attacking a classifier is very different from attacking a detector, and that the structure of detectors - which must search for their own bounding box, and which cannot estimate that box very accurately - likely makes it difficult to make adversarial patterns. Finally, an adversarial pattern on a physical object that could fool a detector would have to be adversarial in the face of a wide family of parametric distortions (scale; view angle; box shift inside the detector; illumination; and so on). Such a pattern would be of great theoretical and practical interest. There is currently no evidence that such patterns exist.

Pulkit Budhiraja, Mark Roman Miller, Abhishek K Modi et al.

Users of Virtual Reality (VR) systems often experience vection, the perception of self-motion in the absence of any physical movement. While vection helps to improve presence in VR, it often leads to a form of motion sickness called cybersickness. Cybersickness is a major deterrent to large scale adoption of VR. Prior work has discovered that changing vection (changing the perceived speed or moving direction) causes more severe cybersickness than steady vection (walking at a constant speed or in a constant direction). Based on this idea, we try to reduce the cybersickness caused by character movements in a First Person Shooter (FPS) game in VR. We propose Rotation Blurring (RB), uniformly blurring the screen during rotational movements to reduce cybersickness. We performed a user study to evaluate the impact of RB in reducing cybersickness. We found that the blurring technique led to an overall reduction in sickness levels of the participants and delayed its onset. Participants who experienced acute levels of cybersickness benefited significantly from this technique.

Jiajun Lu, Hussein Sibai, Evan Fabry et al.

It has been shown that most machine learning algorithms are susceptible to adversarial perturbations. Slightly perturbing an image in a carefully chosen direction in the image space may cause a trained neural network model to misclassify it. Recently, it was shown that physical adversarial examples exist: printing perturbed images then taking pictures of them would still result in misclassification. This raises security and safety concerns. However, these experiments ignore a crucial property of physical objects: the camera can view objects from different distances and at different angles. In this paper, we show experiments that suggest that current constructions of physical adversarial examples do not disrupt object detection from a moving platform. Instead, a trained neural network classifies most of the pictures taken from different distances and angles of a perturbed image correctly. We believe this is because the adversarial property of the perturbation is sensitive to the scale at which the perturbed picture is viewed, so (for example) an autonomous car will misclassify a stop sign only from a small range of distances. Our work raises an important question: can one construct examples that are adversarial for many or most viewing conditions? If so, the construction should offer very significant insights into the internal representation of patterns by deep networks. If not, there is a good prospect that adversarial examples can be reduced to a curiosity with little practical impact.

Jiajun Lu, Theerasit Issaranon, David Forsyth

We describe a method to produce a network where current methods such as DeepFool have great difficulty producing adversarial samples. Our construction suggests some insights into how deep networks work. We provide a reasonable analyses that our construction is difficult to defeat, and show experimentally that our method is hard to defeat with both Type I and Type II attacks using several standard networks and datasets. This SafetyNet architecture is used to an important and novel application SceneProof, which can reliably detect whether an image is a picture of a real scene or not. SceneProof applies to images captured with depth maps (RGBD images) and checks if a pair of image and depth map is consistent. It relies on the relative difficulty of producing naturalistic depth maps for images in post processing. We demonstrate that our SafetyNet is robust to adversarial examples built from currently known attacking approaches.

Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh et al.

Colorization is an ambiguous problem, with multiple viable colorizations for a single grey-level image. However, previous methods only produce the single most probable colorization. Our goal is to model the diversity intrinsic to the problem of colorization and produce multiple colorizations that display long-scale spatial co-ordination. We learn a low dimensional embedding of color fields using a variational autoencoder (VAE). We construct loss terms for the VAE decoder that avoid blurry outputs and take into account the uneven distribution of pixel colors. Finally, we build a conditional model for the multi-modal distribution between grey-level image and the color field embeddings. Samples from this conditional model result in diverse colorization. We demonstrate that our method obtains better diverse colorizations than a standard conditional variational autoencoder (CVAE) model, as well as a recently proposed conditional generative adversarial network (cGAN).

Jason Rock, Theerasit Issaranon, Aditya Deshpande et al.

We show how to extend traditional intrinsic image decompositions to incorporate further layers above albedo and shading. It is hard to obtain data to learn a multi-layer decomposition. Instead, we can learn to decompose an image into layers that are "like this" by authoring generative models for each layer using proxy examples that capture the Platonic ideal (Mondrian images for albedo; rendered 3D primitives for shading; material swatches for shading detail). Our method then generates image layers, one from each model, that explain the image. Our approach rests on innovation in generative models for images. We introduce a Convolutional Variational Auto Encoder (conv-VAE), a novel VAE architecture that can reconstruct high fidelity images. The approach is general, and does not require that layers admit a physical interpretation.