How do multi-modal generative models describe images of recent historical events and figures, whose legacies may be nuanced, multifaceted, or contested? This task necessitates not only accurate visual recognition, but also socio-cultural knowledge and cross-modal reasoning. To address this evaluation challenge, we introduce Century -- a novel dataset of sensitive historical images. This dataset consists of 1,500 images from recent history, created through an automated method combining knowledge graphs and language models with quality and diversity criteria created from the practices of museums and digital archives. We demonstrate through automated and human evaluation that this method produces a set of images that depict events and figures that are diverse across topics and represents all regions of the world. We additionally propose an evaluation framework for evaluating the historical contextualisation capabilities along dimensions of accuracy, thoroughness, and objectivity. We demonstrate this approach by using Century to evaluate four foundation models, scoring performance using both automated and human evaluation. We find that historical contextualisation of sensitive images poses a significant challenge for modern multi-modal foundation models, and offer practical recommendations for how developers can use Century to evaluate improvements to models and applications.

While text-to-image (T2I) generative models have become ubiquitous, they do not necessarily generate images that align with a given prompt. While many metrics and benchmarks have been proposed to evaluate T2I models and alignment metrics, the impact of the evaluation components (prompt sets, human annotations, evaluation task) has not been systematically measured. We find that looking at only *one slice of data*, i.e. one set of capabilities or human annotations, is not enough to obtain stable conclusions that generalise to new conditions or slices when evaluating T2I models or alignment metrics. We address this by introducing an evaluation suite of $>$100K annotations across four human annotation templates that comprehensively evaluates models' capabilities across a range of methods for gathering human annotations and comparing models. In particular, we propose (1) a carefully curated set of prompts -- *Gecko2K*; (2) a statistically grounded method of comparing T2I models; and (3) how to systematically evaluate metrics under three *evaluation tasks* -- *model ordering, pair-wise instance scoring, point-wise instance scoring*. Using this evaluation suite, we evaluate a wide range of metrics and find that a metric may do better in one setting but worse in another. As a result, we introduce a new, interpretable auto-eval metric that is consistently better correlated with human ratings than such existing metrics on our evaluation suite--across different human templates and evaluation settings--and on TIFA160.

The ability to properly benchmark model performance in the face of spurious correlations is important to both build better predictors and increase confidence that models are operating as intended. We demonstrate that characterizing (as opposed to simply quantifying) model mistakes across subgroups is pivotal to properly reflect model biases, which are ignored by standard metrics such as worst-group accuracy or accuracy gap. Inspired by the hypothesis testing framework, we introduce SkewSize, a principled and flexible metric that captures bias from mistakes in a model's predictions. It can be used in multi-class settings or generalised to the open vocabulary setting of generative models. SkewSize is an aggregation of the effect size of the interaction between two categorical variables: the spurious variable representing the bias attribute the model's prediction. We demonstrate the utility of SkewSize in multiple settings including: standard vision models trained on synthetic data, vision models trained on ImageNet, and large scale vision-and-language models from the BLIP-2 family. In each case, the proposed SkewSize is able to highlight biases not captured by other metrics, while also providing insights on the impact of recently proposed techniques, such as instruction tuning.

Text-to-image generative models are capable of producing high-quality images that often faithfully depict concepts described using natural language. In this work, we comprehensively evaluate a range of text-to-image models on numerical reasoning tasks of varying difficulty, and show that even the most advanced models have only rudimentary numerical skills. Specifically, their ability to correctly generate an exact number of objects in an image is limited to small numbers, it is highly dependent on the context the number term appears in, and it deteriorates quickly with each successive number. We also demonstrate that models have poor understanding of linguistic quantifiers (such as “few” or “as many as”), the concept of zero, and struggle with more advanced concepts such as fractional representations. We bundle prompts, generated images and human annotations into GeckoNum, a novel benchmark for evaluation of numerical reasoning.

We propose a new approach to determine correspondences between image pairs in the wild under large changes in illumination, viewpoint, context, and material. While other approaches find correspondences between pairs of images by treating the images independently, we instead condition on both images to implicitly take account of the differences between them. To achieve this, we introduce (i) a spatial attention mechanism (a co-attention module, CoAM) for conditioning the learned features on both images, and (ii) a distinctiveness score used to choose the best matches at test time. CoAM can be added to standard architectures and trained using self-supervision or supervised data, and achieves a significant performance improvement under hard conditions, e.g. large viewpoint changes. We demonstrate that models using CoAM achieve state of the art or competitive results on a wide range of tasks: local matching, camera localization, 3D reconstruction, and image stylization.

Recent work argues that robust training requires substantially larger datasets than those required for standard classification. On CIFAR-10 and CIFAR-100, this translates into a sizable robust-accuracy gap between models trained solely on data from the original training set and those trained with additional data extracted from the "80 Million Tiny Images" dataset (TI-80M). In this paper, we explore how generative models trained solely on the original training set can be leveraged to artificially increase the size of the original training set and improve adversarial robustness to $\ell_p$ norm-bounded perturbations. We identify the sufficient conditions under which incorporating additional generated data can improve robustness, and demonstrate that it is possible to significantly reduce the robust-accuracy gap to models trained with additional real data. Surprisingly, we even show that even the addition of non-realistic random data (generated by Gaussian sampling) can improve robustness. We evaluate our approach on CIFAR-10, CIFAR-100, SVHN and TinyImageNet against $\ell_\infty$ and $\ell_2$ norm-bounded perturbations of size $\epsilon = 8/255$ and $\epsilon = 128/255$, respectively. We show large absolute improvements in robust accuracy compared to previous state-of-the-art methods. Against $\ell_\infty$ norm-bounded perturbations of size $\epsilon = 8/255$, our models achieve 66.10% and 33.49% robust accuracy on CIFAR-10 and CIFAR-100, respectively (improving upon the state-of-the-art by +8.96% and +3.29%). Against $\ell_2$ norm-bounded perturbations of size $\epsilon = 128/255$, our model achieves 78.31% on CIFAR-10 (+3.81%). These results beat most prior works that use external data.

Adversarial training suffers from robust overfitting, a phenomenon where the robust test accuracy starts to decrease during training. In this paper, we focus on reducing robust overfitting by using common data augmentation schemes. We demonstrate that, contrary to previous findings, when combined with model weight averaging, data augmentation can significantly boost robust accuracy. Furthermore, we compare various augmentations techniques and observe that spatial composition techniques work the best for adversarial training. Finally, we evaluate our approach on CIFAR-10 against $\ell_\infty$ and $\ell_2$ norm-bounded perturbations of size $\epsilon = 8/255$ and $\epsilon = 128/255$, respectively. We show large absolute improvements of +2.93% and +2.16% in robust accuracy compared to previous state-of-the-art methods. In particular, against $\ell_\infty$ norm-bounded perturbations of size $\epsilon = 8/255$, our model reaches 60.07% robust accuracy without using any external data. We also achieve a significant performance boost with this approach while using other architectures and datasets such as CIFAR-100, SVHN and TinyImageNet.

View synthesis allows for the generation of new views of a scene given one or more images. This is challenging; it requires comprehensively understanding the 3D scene from images. As a result, current methods typically use multiple images, train on ground-truth depth, or are limited to synthetic data. We propose a novel end-to-end model for this task using a single image at test time; it is trained on real images without any ground-truth 3D information. To this end, we introduce a novel differentiable point cloud renderer that is used to transform a latent 3D point cloud of features into the target view. The projected features are decoded by our refinement network to inpaint missing regions and generate a realistic output image. The 3D component inside of our generative model allows for interpretable manipulation of the latent feature space at test time, e.g. we can animate trajectories from a single image. Additionally, we can generate high resolution images and generalise to other input resolutions. We outperform baselines and prior work on the Matterport, Replica, and RealEstate10K datasets.

Automatic music transcription has primarily focused on transcribing audio to a symbolic music representation (e.g. MIDI or sheet music). However, audio-only approaches often struggle with polyphonic instruments and background noise. In contrast, visual information (e.g. a video of an instrument being played) does not have such ambiguities. In this work, we address the problem of transcribing piano music from visual data alone. We propose an end-to-end deep learning framework that learns to automatically predict note onset events given a video of a person playing the piano. From this, we are able to transcribe the played music in the form of MIDI data. We find that our approach is surprisingly effective in a variety of complex situations, particularly those in which music transcription from audio alone is impossible. We also show that combining audio and video data can improve the transcription obtained from each modality alone.

The objective of this work is to reconstruct the 3D surfaces of sculptures from one or more images using a view-dependent representation. To this end, we train a network, SiDeNet , to predict the Si lhouette and De pth of the surface given a variable number of images; the silhouette is predicted at a different viewpoint from the inputs (e.g. from the side ), while the depth is predicted at the viewpoint of the input images. This has three benefits. First, the network learns a representation of shape beyond that of a single viewpoint, as the silhouette forces it to respect the visual hull, and the depth image forces it to predict concavities (which don’t appear on the visual hull). Second, as the network learns about 3D using the proxy tasks of predicting depth and silhouette images , it is not limited by the resolution of the 3D representation. Finally, using a view-dependent representation (e.g. additionally encoding the viewpoint with the input image) improves the network’s generalisability to unseen objects. Additionally, the network is able to handle the input views in a flexible manner. First, it can ingest a different number of views during training and testing, and it is shown that the reconstruction performance improves as additional views are added at test-time. Second, the additional views do not need to be photometrically consistent. The network is trained and evaluated on two synthetic datasets—a realistic sculpture dataset ( SketchFab ), and ShapeNet. The design of the network is validated by comparing to state of the art methods for a set of tasks. It is shown that (i) passing the input viewpoint (i.e. using a view-dependent representation) improves the network’s generalisability at test time. (ii) Predicting depth/silhouette images allows for higher quality predictions in 2D, as the network is not limited by the chosen latent 3D representation. (iii) On both datasets the method of combining views in a global manner performs better than a local method. Finally, we show that the trained network generalizes to real images, and probe how the network has encoded the latent 3D shape.

Current navigational systems rarely consider generic road landmarks in their navigation instructions, which can lead to mistakes, frustration, and distraction. However, automatic detection of road landmarks is difficult, as current approaches to object detection focus either on out-of-context objects which have special characteristics or on very specific domains. This work presents a future direction for a user-friendly navigational system based on state-of-the-art computer vision techniques that use deep learning for object detection. We propose an automatic hierarchical approach for detecting and classifying a set of static and dynamic road landmarks that would be useful in automatic navigational systems. We further demonstrate a set of optimisations that improve performance and accuracy of the basic system. We evaluate our approach on a natural, 'in-the-wild' dataset to determine how well it handles natural automotive input. Finally, we demonstrate a use-case for our system that extracts information about a vehicle's location and intention.