Thomas Mensink

Thomas Mensink, Jasper Uijlings, Lluis Castrejon et al. · deepmind

We propose Encyclopedic-VQA, a large scale visual question answering (VQA) dataset featuring visual questions about detailed properties of fine-grained categories and instances. It contains 221k unique question+answer pairs each matched with (up to) 5 images, resulting in a total of 1M VQA samples. Moreover, our dataset comes with a controlled knowledge base derived from Wikipedia, marking the evidence to support each answer. Empirically, we show that our dataset poses a hard challenge for large vision+language models as they perform poorly on our dataset: PaLI [14] is state-of-the-art on OK-VQA [37], yet it only achieves 13.0% accuracy on our dataset. Moreover, we experimentally show that progress on answering our encyclopedic questions can be achieved by augmenting large models with a mechanism that retrieves relevant information from the knowledge base. An oracle experiment with perfect retrieval achieves 87.0% accuracy on the single-hop portion of our dataset, and an automatic retrieval-augmented prototype yields 48.8%. We believe that our dataset enables future research on retrieval-augmented vision+language models. It is available at https://github.com/google-research/google-research/tree/master/encyclopedic_vqa .

Mostafa Dehghani, Josip Djolonga, Basil Mustafa et al. · deepmind

The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.

Andrea Agostinelli, Michal Pándy, Jasper Uijlings et al.

Transferability metrics is a maturing field with increasing interest, which aims at providing heuristics for selecting the most suitable source models to transfer to a given target dataset, without fine-tuning them all. However, existing works rely on custom experimental setups which differ across papers, leading to inconsistent conclusions about which transferability metrics work best. In this paper we conduct a large-scale study by systematically constructing a broad range of 715k experimental setup variations. We discover that even small variations to an experimental setup lead to different conclusions about the superiority of a transferability metric over another. Then we propose better evaluations by aggregating across many experiments, enabling to reach more stable conclusions. As a result, we reveal the superiority of LogME at selecting good source datasets to transfer from in a semantic segmentation scenario, NLEEP at selecting good source architectures in an image classification scenario, and GBC at determining which target task benefits most from a given source model. Yet, no single transferability metric works best in all scenarios.

Jasper Uijlings, Thomas Mensink, Vittorio Ferrari

Computer vision is driven by the many datasets available for training or evaluating novel methods. However, each dataset has a different set of class labels, visual definition of classes, images following a specific distribution, annotation protocols, etc. In this paper we explore the automatic discovery of visual-semantic relations between labels across datasets. We aim to understand how instances of a certain class in a dataset relate to the instances of another class in another dataset. Are they in an identity, parent/child, overlap relation? Or is there no link between them at all? To find relations between labels across datasets, we propose methods based on language, on vision, and on their combination. We show that we can effectively discover label relations across datasets, as well as their type. We apply our method to four applications: understand label relations, identify missing aspects, increase label specificity, and predict transfer learning gains. We conclude that label relations cannot be established by looking at the names of classes alone, as they depend strongly on how each of the datasets was constructed.

Lisa Alazraki, Lluis Castrejon, Mostafa Dehghani et al.

This paper studies ensembling in the era of Large Vision-Language Models (LVLMs). Ensembling is a classical method to combine different models to get increased performance. In the recent work on Encyclopedic-VQA the authors examine a wide variety of models to solve their task: from vanilla LVLMs, to models including the caption as extra context, to models augmented with Lens-based retrieval of Wikipedia pages. Intuitively these models are highly complementary, which should make them ideal for ensembling. Indeed, an oracle experiment shows potential gains from 48.8% accuracy (the best single model) all the way up to 67% (best possible ensemble). So it is a trivial exercise to create an ensemble with substantial real gains. Or is it?

Emiel Hoogeboom, David Ruhe, Jonathan Heek et al.

It is currently difficult to distill discrete diffusion models. In contrast, continuous diffusion literature has many distillation approaches methods that can reduce sampling steps to a handful. Our method, Discrete Moment Matching Distillation (D-MMD), leverages ideas that have been highly successful in the continuous domain. Whereas previous discrete distillation methods collapse, D-MMD maintains high quality and diversity (given sufficient sampling steps). This is demonstrated on both text and image datasets. Moreover, the newly distilled generators can outperform their teachers.

Jonathan Heek, Emiel Hoogeboom, Thomas Mensink et al.

We present Unified Latents (UL), a framework for learning latent representations that are jointly regularized by a diffusion prior and decoded by a diffusion model. By linking the encoder's output noise to the prior's minimum noise level, we obtain a simple training objective that provides a tight upper bound on the latent bitrate. On ImageNet-512, our approach achieves competitive FID of 1.4, with high reconstruction quality (PSNR) while requiring fewer training FLOPs than models trained on Stable Diffusion latents. On Kinetics-600, we set a new state-of-the-art FVD of 1.3.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Grigory Bartosh, David Ruhe, Emiel Hoogeboom et al.

Diffusion models achieve state-of-the-art generative performance but suffer from high computational costs during inference due to the repeated evaluation of a heavy neural network. In this work, we propose Dual-Rate Diffusion, a method to accelerate sampling by interleaving the execution of a heavy high-capacity context encoder and a light efficient denoising model. The context encoder is evaluated sparsely to extract high-dimensional features, which are effectively reused by the light denoising model at every step to refine the sample efficiently. This approach significantly accelerates inference without compromising sample quality. On ImageNet benchmarks, Dual-Rate Diffusion matches the performance of standard baselines while reducing computational cost by a factor of $2$-$4$. Furthermore, we demonstrate that our method is compatible with distillation techniques, such as Moment Matching Distillation, enabling further efficiency gains in few-step generation.

Imagen-Team-Google, Jason Baldridge, Jakob Bauer et al.

We introduce Imagen 3, a latent diffusion model that generates high quality images from text prompts. We describe our quality and responsibility evaluations. Imagen 3 is preferred over other state-of-the-art (SOTA) models at the time of evaluation. In addition, we discuss issues around safety and representation, as well as methods we used to minimize the potential harm of our models.

Kartik Gupta, Amir Rahimi, Thalaiyasingam Ajanthan et al.

Calibrating neural networks is of utmost importance when employing them in safety-critical applications where the downstream decision making depends on the predicted probabilities. Measuring calibration error amounts to comparing two empirical distributions. In this work, we introduce a binning-free calibration measure inspired by the classical Kolmogorov-Smirnov (KS) statistical test in which the main idea is to compare the respective cumulative probability distributions. From this, by approximating the empirical cumulative distribution using a differentiable function via splines, we obtain a recalibration function, which maps the network outputs to actual (calibrated) class assignment probabilities. The spine-fitting is performed using a held-out calibration set and the obtained recalibration function is evaluated on an unseen test set. We tested our method against existing calibration approaches on various image classification datasets and our spline-based recalibration approach consistently outperforms existing methods on KS error as well as other commonly used calibration measures. Our Code is available at https://github.com/kartikgupta-at-anu/spline-calibration.

Hoàng-Ân Lê, Tushar Nimbhorkar, Thomas Mensink et al.

There hardly exists any large-scale datasets with dense optical flow of non-rigid motion from real-world imagery as of today. The reason lies mainly in the required setup to derive ground truth optical flows: a series of images with known camera poses along its trajectory, and an accurate 3D model from a textured scene. Human annotation is not only too tedious for large databases, it can simply hardly contribute to accurate optical flow. To circumvent the need for manual annotation, we propose a framework to automatically generate optical flow from real-world videos. The method extracts and matches objects from video frames to compute initial constraints, and applies a deformation over the objects of interest to obtain dense optical flow fields. We propose several ways to augment the optical flow variations. Extensive experimental results show that training on our automatically generated optical flow outperforms methods that are trained on rigid synthetic data using FlowNet-S, LiteFlowNet, PWC-Net, and RAFT. Datasets and implementation of our optical flow generation framework are released at https://github.com/lhoangan/arap_flow

Emiel Hoogeboom, Thomas Mensink, Jonathan Heek et al.

Latent diffusion models have become the popular choice for scaling up diffusion models for high resolution image synthesis. Compared to pixel-space models that are trained end-to-end, latent models are perceived to be more efficient and to produce higher image quality at high resolution. Here we challenge these notions, and show that pixel-space models can be very competitive to latent models both in quality and efficiency, achieving 1.5 FID on ImageNet512 and new SOTA results on ImageNet128, ImageNet256 and Kinetics600. We present a simple recipe for scaling end-to-end pixel-space diffusion models to high resolutions. 1: Use the sigmoid loss-weighting (Kingma & Gao, 2023) with our prescribed hyper-parameters. 2: Use our simplified memory-efficient architecture with fewer skip-connections. 3: Scale the model to favor processing the image at a high resolution with fewer parameters, rather than using more parameters at a lower resolution. Combining these with guidance intervals, we obtain a family of pixel-space diffusion models we call Simpler Diffusion (SiD2).

Lluis Castrejon, Thomas Mensink, Howard Zhou et al.

Combining Large Language Models (LLMs) with external specialized tools (LLMs+tools) is a recent paradigm to solve multimodal tasks such as Visual Question Answering (VQA). While this approach was demonstrated to work well when optimized and evaluated for each individual benchmark, in practice it is crucial for the next generation of real-world AI systems to handle a broad range of multimodal problems. Therefore we pose the VQA problem from a unified perspective and evaluate a single system on a varied suite of VQA tasks including counting, spatial reasoning, OCR-based reasoning, visual pointing, external knowledge, and more. In this setting, we demonstrate that naively applying the LLM+tools approach using the combined set of all tools leads to poor results. This motivates us to introduce HAMMR: HierArchical MultiModal React. We start from a multimodal ReAct-based system and make it hierarchical by enabling our HAMMR agents to call upon other specialized agents. This enhances the compositionality of the LLM+tools approach, which we show to be critical for obtaining high accuracy on generic VQA. Concretely, on our generic VQA suite, HAMMR outperforms the naive LLM+tools approach by 19.5%. Additionally, HAMMR achieves state-of-the-art results on this task, outperforming the generic standalone PaLI-X VQA model by 5.0%.

Tim Salimans, Thomas Mensink, Jonathan Heek et al.

We present a new method for making diffusion models faster to sample. The method distills many-step diffusion models into few-step models by matching conditional expectations of the clean data given noisy data along the sampling trajectory. Our approach extends recently proposed one-step methods to the multi-step case, and provides a new perspective by interpreting these approaches in terms of moment matching. By using up to 8 sampling steps, we obtain distilled models that outperform not only their one-step versions but also their original many-step teacher models, obtaining new state-of-the-art results on the Imagenet dataset. We also show promising results on a large text-to-image model where we achieve fast generation of high resolution images directly in image space, without needing autoencoders or upsamplers.

Thomas Mensink, Pascal Mettes

Mixup is a widely adopted strategy for training deep networks, where additional samples are augmented by interpolating inputs and labels of training pairs. Mixup has shown to improve classification performance, network calibration, and out-of-distribution generalisation. While effective, a cornerstone of Mixup, namely that networks learn linear behaviour patterns between classes, is only indirectly enforced since the output interpolation is performed at the probability level. This paper seeks to address this limitation by mixing the classifiers directly instead of mixing the labels for each mixed pair. We propose to define the target of each augmented sample as a uniquely new classifier, whose parameters are a linear interpolation of the classifier vectors of the input pair. The space of all possible classifiers is continuous and spans all interpolations between classifier pairs. To make optimisation tractable, we propose a dual-contrastive Infinite Class Mixup loss, where we contrast the classifier of a mixed pair to both the classifiers and the predicted outputs of other mixed pairs in a batch. Infinite Class Mixup is generic in nature and applies to many variants of Mixup. Empirically, we show that it outperforms standard Mixup and variants such as RegMixup and Remix on balanced, long-tailed, and data-constrained benchmarks, highlighting its broad applicability.

Andrea Agostinelli, Jasper Uijlings, Thomas Mensink et al.

We address the problem of ensemble selection in transfer learning: Given a large pool of source models we want to select an ensemble of models which, after fine-tuning on the target training set, yields the best performance on the target test set. Since fine-tuning all possible ensembles is computationally prohibitive, we aim at predicting performance on the target dataset using a computationally efficient transferability metric. We propose several new transferability metrics designed for this task and evaluate them in a challenging and realistic transfer learning setup for semantic segmentation: we create a large and diverse pool of source models by considering 17 source datasets covering a wide variety of image domain, two different architectures, and two pre-training schemes. Given this pool, we then automatically select a subset to form an ensemble performing well on a given target dataset. We compare the ensemble selected by our method to two baselines which select a single source model, either (1) from the same pool as our method; or (2) from a pool containing large source models, each with similar capacity as an ensemble. Averaged over 17 target datasets, we outperform these baselines by 6.0% and 2.5% relative mean IoU, respectively.

Michal Pándy, Andrea Agostinelli, Jasper Uijlings et al.

Transfer learning has become a popular method for leveraging pre-trained models in computer vision. However, without performing computationally expensive fine-tuning, it is difficult to quantify which pre-trained source models are suitable for a specific target task, or, conversely, to which tasks a pre-trained source model can be easily adapted to. In this work, we propose Gaussian Bhattacharyya Coefficient (GBC), a novel method for quantifying transferability between a source model and a target dataset. In a first step we embed all target images in the feature space defined by the source model, and represent them with per-class Gaussians. Then, we estimate their pairwise class separability using the Bhattacharyya coefficient, yielding a simple and effective measure of how well the source model transfers to the target task. We evaluate GBC on image classification tasks in the context of dataset and architecture selection. Further, we also perform experiments on the more complex semantic segmentation transferability estimation task. We demonstrate that GBC outperforms state-of-the-art transferability metrics on most evaluation criteria in the semantic segmentation settings, matches the performance of top methods for dataset transferability in image classification, and performs best on architecture selection problems for image classification.

Thomas Mensink, Jasper Uijlings, Alina Kuznetsova et al.

Transfer learning enables to re-use knowledge learned on a source task to help learning a target task. A simple form of transfer learning is common in current state-of-the-art computer vision models, i.e. pre-training a model for image classification on the ILSVRC dataset, and then fine-tune on any target task. However, previous systematic studies of transfer learning have been limited and the circumstances in which it is expected to work are not fully understood. In this paper we carry out an extensive experimental exploration of transfer learning across vastly different image domains (consumer photos, autonomous driving, aerial imagery, underwater, indoor scenes, synthetic, close-ups) and task types (semantic segmentation, object detection, depth estimation, keypoint detection). Importantly, these are all complex, structured output tasks types relevant to modern computer vision applications. In total we carry out over 2000 transfer learning experiments, including many where the source and target come from different image domains, task types, or both. We systematically analyze these experiments to understand the impact of image domain, task type, and dataset size on transfer learning performance. Our study leads to several insights and concrete recommendations: (1) for most tasks there exists a source which significantly outperforms ILSVRC'12 pre-training; (2) the image domain is the most important factor for achieving positive transfer; (3) the source dataset should \emph{include} the image domain of the target dataset to achieve best results; (4) at the same time, we observe only small negative effects when the image domain of the source task is much broader than that of the target; (5) transfer across task types can be beneficial, but its success is heavily dependent on both the source and target task types.

Hoang-An Le, Thomas Mensink, Partha Das et al.

Multimodal large-scale datasets for outdoor scenes are mostly designed for urban driving problems. The scenes are highly structured and semantically different from scenarios seen in nature-centered scenes such as gardens or parks. To promote machine learning methods for nature-oriented applications, such as agriculture and gardening, we propose the multimodal synthetic dataset for Enclosed garDEN scenes (EDEN). The dataset features more than 300K images captured from more than 100 garden models. Each image is annotated with various low/high-level vision modalities, including semantic segmentation, depth, surface normals, intrinsic colors, and optical flow. Experimental results on the state-of-the-art methods for semantic segmentation and monocular depth prediction, two important tasks in computer vision, show positive impact of pre-training deep networks on our dataset for unstructured natural scenes. The dataset and related materials will be available at https://lhoangan.github.io/eden.

Hoang-An Le, Thomas Mensink, Partha Das et al.

In this paper the argument is made that for true novel view synthesis of objects, where the object can be synthesized from any viewpoint, an explicit 3D shape representation isdesired. Our method estimates point clouds to capture the geometry of the object, which can be freely rotated into the desired view and then projected into a new image. This image, however, is sparse by nature and hence this coarse view is used as the input of an image completion network to obtain the dense target view. The point cloud is obtained using the predicted pixel-wise depth map, estimated from a single RGB input image,combined with the camera intrinsics. By using forward warping and backward warpingbetween the input view and the target view, the network can be trained end-to-end without supervision on depth. The benefit of using point clouds as an explicit 3D shape for novel view synthesis is experimentally validated on the 3D ShapeNet benchmark. Source code and data will be available at https://lhoangan.github.io/pc4novis/.

Rick Groenendijk, Sezer Karaoglu, Theo Gevers et al.

Many interesting tasks in machine learning and computer vision are learned by optimising an objective function defined as a weighted linear combination of multiple losses. The final performance is sensitive to choosing the correct (relative) weights for these losses. Finding a good set of weights is often done by adopting them into the set of hyper-parameters, which are set using an extensive grid search. This is computationally expensive. In this paper, we propose a weighting scheme based on the coefficient of variations and set the weights based on properties observed while training the model. The proposed method incorporates a measure of uncertainty to balance the losses, and as a result the loss weights evolve during training without requiring another (learning based) optimisation. In contrast to many loss weighting methods in literature, we focus on single-task multi-loss problems, such as monocular depth estimation and semantic segmentation, and show that multi-task approaches for loss weighting do not work on those single-tasks. The validity of the approach is shown empirically for depth estimation and semantic segmentation on multiple datasets.

Yunlu Chen, Vincent Tao Hu, Efstratios Gavves et al.

This paper introduces data augmentation for point clouds by interpolation between examples. Data augmentation by interpolation has shown to be a simple and effective approach in the image domain. Such a mixup is however not directly transferable to point clouds, as we do not have a one-to-one correspondence between the points of two different objects. In this paper, we define data augmentation between point clouds as a shortest path linear interpolation. To that end, we introduce PointMixup, an interpolation method that generates new examples through an optimal assignment of the path function between two point clouds. We prove that our PointMixup finds the shortest path between two point clouds and that the interpolation is assignment invariant and linear. With the definition of interpolation, PointMixup allows to introduce strong interpolation-based regularizers such as mixup and manifold mixup to the point cloud domain. Experimentally, we show the potential of PointMixup for point cloud classification, especially when examples are scarce, as well as increased robustness to noise and geometric transformations to points. The code for PointMixup and the experimental details are publicly available.

Amir Rahimi, Thomas Mensink, Kartik Gupta et al.

Calibration of neural networks is a critical aspect to consider when incorporating machine learning models in real-world decision-making systems where the confidence of decisions are equally important as the decisions themselves. In recent years, there is a surge of research on neural network calibration and the majority of the works can be categorized into post-hoc calibration methods, defined as methods that learn an additional function to calibrate an already trained base network. In this work, we intend to understand the post-hoc calibration methods from a theoretical point of view. Especially, it is known that minimizing Negative Log-Likelihood (NLL) will lead to a calibrated network on the training set if the global optimum is attained (Bishop, 1994). Nevertheless, it is not clear learning an additional function in a post-hoc manner would lead to calibration in the theoretical sense. To this end, we prove that even though the base network ($f$) does not lead to the global optimum of NLL, by adding additional layers ($g$) and minimizing NLL by optimizing the parameters of $g$ one can obtain a calibrated network $g \circ f$. This not only provides a less stringent condition to obtain a calibrated network but also provides a theoretical justification of post-hoc calibration methods. Our experiments on various image classification benchmarks confirm the theory.

Alex Bewley, Pei Sun, Thomas Mensink et al.

This paper presents a novel 3D object detection framework that processes LiDAR data directly on its native representation: range images. Benefiting from the compactness of range images, 2D convolutions can efficiently process dense LiDAR data of a scene. To overcome scale sensitivity in this perspective view, a novel range-conditioned dilation (RCD) layer is proposed to dynamically adjust a continuous dilation rate as a function of the measured range. Furthermore, localized soft range gating combined with a 3D box-refinement stage improves robustness in occluded areas, and produces overall more accurate bounding box predictions. On the public large-scale Waymo Open Dataset, our method sets a new baseline for range-based 3D detection, outperforming multiview and voxel-based methods over all ranges with unparalleled performance at long range detection.

Rick Groenendijk, Sezer Karaoglu, Theo Gevers et al.

In this paper we address the benefit of adding adversarial training to the task of monocular depth estimation. A model can be trained in a self-supervised setting on stereo pairs of images, where depth (disparities) are an intermediate result in a right-to-left image reconstruction pipeline. For the quality of the image reconstruction and disparity prediction, a combination of different losses is used, including L1 image reconstruction losses and left-right disparity smoothness. These are local pixel-wise losses, while depth prediction requires global consistency. Therefore, we extend the self-supervised network to become a Generative Adversarial Network (GAN), by including a discriminator which should tell apart reconstructed (fake) images from real images. We evaluate Vanilla GANs, LSGANs and Wasserstein GANs in combination with different pixel-wise reconstruction losses. Based on extensive experimental evaluation, we conclude that adversarial training is beneficial if and only if the reconstruction loss is not too constrained. Even though adversarial training seems promising because it promotes global consistency, non-adversarial training outperforms (or is on par with) any method trained with a GAN when a constrained reconstruction loss is used in combination with batch normalisation. Based on the insights of our experimental evaluation we obtain state-of-the art monocular depth estimation results by using batch normalisation and different output scales.

Yunlu Chen, Thomas Mensink, Efstratios Gavves

A key challenge for RGB-D segmentation is how to effectively incorporate 3D geometric information from the depth channel into 2D appearance features. We propose to model the effective receptive field of 2D convolution based on the scale and locality from the 3D neighborhood. Standard convolutions are local in the image space ($u, v$), often with a fixed receptive field of 3x3 pixels. We propose to define convolutions local with respect to the corresponding point in the 3D real-world space ($x, y, z$), where the depth channel is used to adapt the receptive field of the convolution, which yields the resulting filters invariant to scale and focusing on the certain range of depth. We introduce 3D Neighborhood Convolution (3DN-Conv), a convolutional operator around 3D neighborhoods. Further, we can use estimated depth to use our RGB-D based semantic segmentation model from RGB input. Experimental results validate that our proposed 3DN-Conv operator improves semantic segmentation, using either ground-truth depth (RGB-D) or estimated depth (RGB).

Hoang-An Le, Anil S. Baslamisli, Thomas Mensink et al.

Optical flow, semantic segmentation, and surface normals represent different information modalities, yet together they bring better cues for scene understanding problems. In this paper, we study the influence between the three modalities: how one impacts on the others and their efficiency in combination. We employ a modular approach using a convolutional refinement network which is trained supervised but isolated from RGB images to enforce joint modality features. To assist the training process, we create a large-scale synthetic outdoor dataset that supports dense annotation of semantic segmentation, optical flow, and surface normals. The experimental results show positive influence among the three modalities, especially for objects' boundaries, region consistency, and scene structures.

Ysbrand Galama, Thomas Mensink

We are interested in learning visual representations which allow for 3D manipulations of visual objects based on a single 2D image. We cast this into an image-to-image transformation task, and propose Iterative Generative Adversarial Networks (IterGANs) which iteratively transform an input image into an output image. Our models learn a visual representation that can be used for objects seen in training, but also for never seen objects. Since object manipulation requires a full understanding of the geometry and appearance of the object, our IterGANs learn an implicit 3D model and a full appearance model of the object, which are both inferred from a single (test) image. Two advantages of IterGANs are that the intermediate generated images can be used for an additional supervision signal, even in an unsupervised fashion, and that the number of iterations can be used as a control signal to steer the transformation. Experiments on rotated objects and scenes show how IterGANs help with the generation process.

Spencer Cappallo, Stacey Svetlichnaya, Pierre Garrigues et al.

Over the past decade, emoji have emerged as a new and widespread form of digital communication, spanning diverse social networks and spoken languages. We propose to treat these ideograms as a new modality in their own right, distinct in their semantic structure from both the text in which they are often embedded as well as the images which they resemble. As a new modality, emoji present rich novel possibilities for representation and interaction. In this paper, we explore the challenges that arise naturally from considering the emoji modality through the lens of multimedia research. Specifically, the ways in which emoji can be related to other common modalities such as text and images. To do so, we first present a large scale dataset of real-world emoji usage collected from Twitter. This dataset contains examples of both text-emoji and image-emoji relationships. We present baseline results on the challenge of predicting emoji from both text and images, using state-of-the-art neural networks. Further, we offer a first consideration into the problem of how to account for new, unseen emoji - a relevant issue as the emoji vocabulary continues to expand on a yearly basis. Finally, we present results for multimedia retrieval using emoji as queries.

Spencer Cappallo, Thomas Mensink, Cees G. M. Snoek

Retrieval of live, user-broadcast video streams is an under-addressed and increasingly relevant challenge. The on-line nature of the problem requires temporal evaluation and the unforeseeable scope of potential queries motivates an approach which can accommodate arbitrary search queries. To account for the breadth of possible queries, we adopt a no-example approach to query retrieval, which uses a query's semantic relatedness to pre-trained concept classifiers. To adapt to shifting video content, we propose memory pooling and memory welling methods that favor recent information over long past content. We identify two stream retrieval tasks, instantaneous retrieval at any particular time and continuous retrieval over a prolonged duration, and propose means for evaluating them. Three large scale video datasets are adapted to the challenge of stream retrieval. We report results for our search methods on the new stream retrieval tasks, as well as demonstrate their efficacy in a traditional, non-streaming video task.

Rocco De Rosa, Thomas Mensink, Barbara Caputo

As we enter into the big data age and an avalanche of images have become readily available, recognition systems face the need to move from close, lab settings where the number of classes and training data are fixed, to dynamic scenarios where the number of categories to be recognized grows continuously over time, as well as new data providing useful information to update the system. Recent attempts, like the open world recognition framework, tried to inject dynamics into the system by detecting new unknown classes and adding them incrementally, while at the same time continuously updating the models for the known classes. incrementally adding new classes and detecting instances from unknown classes, while at the same time continuously updating the models for the known classes. In this paper we argue that to properly capture the intrinsic dynamic of open world recognition, it is necessary to add to these aspects (a) the incremental learning of the underlying metric, (b) the incremental estimate of confidence thresholds for the unknown classes, and (c) the use of local learning to precisely describe the space of classes. We extend three existing metric learning algorithms towards these goals by using online metric learning. Experimentally we validate our approach on two large-scale datasets in different learning scenarios. For all these scenarios our proposed methods outperform their non-online counterparts. We conclude that local and online learning is important to capture the full dynamics of open world recognition.

Amirhossein Habibian, Thomas Mensink, Cees G. M. Snoek

This paper aims for event recognition when video examples are scarce or even completely absent. The key in such a challenging setting is a semantic video representation. Rather than building the representation from individual attribute detectors and their annotations, we propose to learn the entire representation from freely available web videos and their descriptions using an embedding between video features and term vectors. In our proposed embedding, which we call VideoStory, the correlations between the terms are utilized to learn a more effective representation by optimizing a joint objective balancing descriptiveness and predictability.We show how learning the VideoStory using a multimodal predictability loss, including appearance, motion and audio features, results in a better predictable representation. We also propose a variant of VideoStory to recognize an event in video from just the important terms in a text query by introducing a term sensitive descriptiveness loss. Our experiments on three challenging collections of web videos from the NIST TRECVID Multimedia Event Detection and Columbia Consumer Videos datasets demonstrate: i) the advantages of VideoStory over representations using attributes or alternative embeddings, ii) the benefit of fusing video modalities by an embedding over common strategies, iii) the complementarity of term sensitive descriptiveness and multimodal predictability for event recognition without examples. By it abilities to improve predictability upon any underlying video feature while at the same time maximizing semantic descriptiveness, VideoStory leads to state-of-the-art accuracy for both few- and zero-example recognition of events in video.

Mihir Jain, Jan C. van Gemert, Thomas Mensink et al.

The goal of this paper is to recognize actions in video without the need for examples. Different from traditional zero-shot approaches we do not demand the design and specification of attribute classifiers and class-to-attribute mappings to allow for transfer from seen classes to unseen classes. Our key contribution is objects2action, a semantic word embedding that is spanned by a skip-gram model of thousands of object categories. Action labels are assigned to an object encoding of unseen video based on a convex combination of action and object affinities. Our semantic embedding has three main characteristics to accommodate for the specifics of actions. First, we propose a mechanism to exploit multiple-word descriptions of actions and objects. Second, we incorporate the automated selection of the most responsive objects per action. And finally, we demonstrate how to extend our zero-shot approach to the spatio-temporal localization of actions in video. Experiments on four action datasets demonstrate the potential of our approach.

Efstratios Gavves, Thomas Mensink, Tatiana Tommasi et al.

How can we reuse existing knowledge, in the form of available datasets, when solving a new and apparently unrelated target task from a set of unlabeled data? In this work we make a first contribution to answer this question in the context of image classification. We frame this quest as an active learning problem and use zero-shot classifiers to guide the learning process by linking the new task to the existing classifiers. By revisiting the dual formulation of adaptive SVM, we reveal two basic conditions to choose greedily only the most relevant samples to be annotated. On this basis we propose an effective active learning algorithm which learns the best possible target classification model with minimum human labeling effort. Extensive experiments on two challenging datasets show the value of our approach compared to the state-of-the-art active learning methodologies, as well as its potential to reuse past datasets with minimal effort for future tasks.