Jasper Uijlings

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 .

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?

Jasper Uijlings, Xingyi Zhou, Xiuye Gu et al.

Understanding objects in videos in terms of fine-grained localization masks and detailed semantic properties is a fundamental task in video understanding. In this paper, we propose VoCap, a flexible video model that consumes a video and a prompt of various modalities (text, box or mask), and produces a spatio-temporal masklet with a corresponding object-centric caption. As such our model addresses simultaneously the tasks of promptable video object segmentation, referring expression segmentation, and object captioning. Since obtaining data for this task is tedious and expensive, we propose to annotate an existing large-scale segmentation dataset (SAV) with pseudo object captions. We do so by preprocessing videos with their ground-truth masks to highlight the object of interest and feed this to a large Vision Language Model (VLM). For an unbiased evaluation, we collect manual annotations on the validation set. We call the resulting dataset SAV-Caption. We train our VoCap model at scale on a SAV-Caption together with a mix of other image and video datasets. Our model yields state-of-the-art results on referring expression video object segmentation, is competitive on semi-supervised video object segmentation, and establishes a benchmark for video object captioning. Our dataset will be made available at https://github.com/google-deepmind/vocap.

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%.

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.

Michael Gygli, Jasper Uijlings, Vittorio Ferrari

This paper proposes to make a first step towards compatible and hence reusable network components. Rather than training networks for different tasks independently, we adapt the training process to produce network components that are compatible across tasks. In particular, we split a network into two components, a features extractor and a target task head, and propose various approaches to accomplish compatibility between them. We systematically analyse these approaches on the task of image classification on standard datasets. We demonstrate that we can produce components which are directly compatible without any fine-tuning or compromising accuracy on the original tasks. Afterwards, we demonstrate the use of compatible components on three applications: Unsupervised domain adaptation, transferring classifiers across feature extractors with different architectures, and increasing the computational efficiency of transfer learning.

Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo et al.

We propose Localized Narratives, a new form of multimodal image annotations connecting vision and language. We ask annotators to describe an image with their voice while simultaneously hovering their mouse over the region they are describing. Since the voice and the mouse pointer are synchronized, we can localize every single word in the description. This dense visual grounding takes the form of a mouse trace segment per word and is unique to our data. We annotated 849k images with Localized Narratives: the whole COCO, Flickr30k, and ADE20K datasets, and 671k images of Open Images, all of which we make publicly available. We provide an extensive analysis of these annotations showing they are diverse, accurate, and efficient to produce. We also demonstrate their utility on the application of controlled image captioning.

Theodora Kontogianni, Michael Gygli, Jasper Uijlings et al.

In interactive object segmentation a user collaborates with a computer vision model to segment an object. Recent works employ convolutional neural networks for this task: Given an image and a set of corrections made by the user as input, they output a segmentation mask. These approaches achieve strong performance by training on large datasets but they keep the model parameters unchanged at test time. Instead, we recognize that user corrections can serve as sparse training examples and we propose a method that capitalizes on that idea to update the model parameters on-the-fly to the data at hand. Our approach enables the adaptation to a particular object and its background, to distributions shifts in a test set, to specific object classes, and even to large domain changes, where the imaging modality changes between training and testing. We perform extensive experiments on 8 diverse datasets and show: Compared to a model with frozen parameters, our method reduces the required corrections (i) by 9%-30% when distribution shifts are small between training and testing; (ii) by 12%-44% when specializing to a specific class; (iii) and by 60% and 77% when we completely change domain between training and testing.

Alina Kuznetsova, Hassan Rom, Neil Alldrin et al.

We present Open Images V4, a dataset of 9.2M images with unified annotations for image classification, object detection and visual relationship detection. The images have a Creative Commons Attribution license that allows to share and adapt the material, and they have been collected from Flickr without a predefined list of class names or tags, leading to natural class statistics and avoiding an initial design bias. Open Images V4 offers large scale across several dimensions: 30.1M image-level labels for 19.8k concepts, 15.4M bounding boxes for 600 object classes, and 375k visual relationship annotations involving 57 classes. For object detection in particular, we provide 15x more bounding boxes than the next largest datasets (15.4M boxes on 1.9M images). The images often show complex scenes with several objects (8 annotated objects per image on average). We annotated visual relationships between them, which support visual relationship detection, an emerging task that requires structured reasoning. We provide in-depth comprehensive statistics about the dataset, we validate the quality of the annotations, we study how the performance of several modern models evolves with increasing amounts of training data, and we demonstrate two applications made possible by having unified annotations of multiple types coexisting in the same images. We hope that the scale, quality, and variety of Open Images V4 will foster further research and innovation even beyond the areas of image classification, object detection, and visual relationship detection.

Ksenia Konyushkova, Jasper Uijlings, Christoph Lampert et al.

We introduce Intelligent Annotation Dialogs for bounding box annotation. We train an agent to automatically choose a sequence of actions for a human annotator to produce a bounding box in a minimal amount of time. Specifically, we consider two actions: box verification, where the annotator verifies a box generated by an object detector, and manual box drawing. We explore two kinds of agents, one based on predicting the probability that a box will be positively verified, and the other based on reinforcement learning. We demonstrate that (1) our agents are able to learn efficient annotation strategies in several scenarios, automatically adapting to the image difficulty, the desired quality of the boxes, and the detector strength; (2) in all scenarios the resulting annotation dialogs speed up annotation compared to manual box drawing alone and box verification alone, while also outperforming any fixed combination of verification and drawing in most scenarios; (3) in a realistic scenario where the detector is iteratively re-trained, our agents evolve a series of strategies that reflect the shifting trade-off between verification and drawing as the detector grows stronger.

Jasper Uijlings, Stefan Popov, Vittorio Ferrari

We propose to revisit knowledge transfer for training object detectors on target classes from weakly supervised training images, helped by a set of source classes with bounding-box annotations. We present a unified knowledge transfer framework based on training a single neural network multi-class object detector over all source classes, organized in a semantic hierarchy. This generates proposals with scores at multiple levels in the hierarchy, which we use to explore knowledge transfer over a broad range of generality, ranging from class-specific (bicycle to motorbike) to class-generic (objectness to any class). Experiments on the 200 object classes in the ILSVRC 2013 detection dataset show that our technique: (1) leads to much better performance on the target classes (70.3% CorLoc, 36.9% mAP) than a weakly supervised baseline which uses manually engineered objectness [11] (50.5% CorLoc, 25.4% mAP). (2) delivers target object detectors reaching 80% of the mAP of their fully supervised counterparts. (3) outperforms the best reported transfer learning results on this dataset (+41% CorLoc and +3% mAP over [18, 46], +16.2% mAP over [32]). Moreover, we also carry out several across-dataset knowledge transfer experiments [27, 24, 35] and find that (4) our technique outperforms the weakly supervised baseline in all dataset pairs by 1.5x-1.9x, establishing its general applicability.

Zbigniew Wojna, Vittorio Ferrari, Sergio Guadarrama et al.

Many machine vision applications, such as semantic segmentation and depth prediction, require predictions for every pixel of the input image. Models for such problems usually consist of encoders which decrease spatial resolution while learning a high-dimensional representation, followed by decoders who recover the original input resolution and result in low-dimensional predictions. While encoders have been studied rigorously, relatively few studies address the decoder side. This paper presents an extensive comparison of a variety of decoders for a variety of pixel-wise tasks ranging from classification, regression to synthesis. Our contributions are: (1) Decoders matter: we observe significant variance in results between different types of decoders on various problems. (2) We introduce new residual-like connections for decoders. (3) We introduce a novel decoder: bilinear additive upsampling. (4) We explore prediction artifacts.

Holger Caesar, Jasper Uijlings, Vittorio Ferrari

Semantic classes can be either things (objects with a well-defined shape, e.g. car, person) or stuff (amorphous background regions, e.g. grass, sky). While lots of classification and detection works focus on thing classes, less attention has been given to stuff classes. Nonetheless, stuff classes are important as they allow to explain important aspects of an image, including (1) scene type; (2) which thing classes are likely to be present and their location (through contextual reasoning); (3) physical attributes, material types and geometric properties of the scene. To understand stuff and things in context we introduce COCO-Stuff, which augments all 164K images of the COCO 2017 dataset with pixel-wise annotations for 91 stuff classes. We introduce an efficient stuff annotation protocol based on superpixels, which leverages the original thing annotations. We quantify the speed versus quality trade-off of our protocol and explore the relation between annotation time and boundary complexity. Furthermore, we use COCO-Stuff to analyze: (a) the importance of stuff and thing classes in terms of their surface cover and how frequently they are mentioned in image captions; (b) the spatial relations between stuff and things, highlighting the rich contextual relations that make our dataset unique; (c) the performance of a modern semantic segmentation method on stuff and thing classes, and whether stuff is easier to segment than things.

Holger Caesar, Jasper Uijlings, Vittorio Ferrari

We propose a novel method for semantic segmentation, the task of labeling each pixel in an image with a semantic class. Our method combines the advantages of the two main competing paradigms. Methods based on region classification offer proper spatial support for appearance measurements, but typically operate in two separate stages, none of which targets pixel labeling performance at the end of the pipeline. More recent fully convolutional methods are capable of end-to-end training for the final pixel labeling, but resort to fixed patches as spatial support. We show how to modify modern region-based approaches to enable end-to-end training for semantic segmentation. This is achieved via a differentiable region-to-pixel layer and a differentiable free-form Region-of-Interest pooling layer. Our method improves the state-of-the-art in terms of class-average accuracy with 64.0% on SIFT Flow and 49.9% on PASCAL Context, and is particularly accurate at object boundaries.

Holger Caesar, Jasper Uijlings, Vittorio Ferrari

Semantic segmentation is the task of assigning a class-label to each pixel in an image. We propose a region-based semantic segmentation framework which handles both full and weak supervision, and addresses three common problems: (1) Objects occur at multiple scales and therefore we should use regions at multiple scales. However, these regions are overlapping which creates conflicting class predictions at the pixel-level. (2) Class frequencies are highly imbalanced in realistic datasets. (3) Each pixel can only be assigned to a single class, which creates competition between classes. We address all three problems with a joint calibration method which optimizes a multi-class loss defined over the final pixel-level output labeling, as opposed to simply region classification. Our method outperforms the state-of-the-art on the popular SIFT Flow [18] dataset in both the fully and weakly supervised setting by a considerably margin (+6% and +10%, respectively).

Jasper Uijlings, Vittorio Ferrari

Intuitively, the appearance of true object boundaries varies from image to image. Hence the usual monolithic approach of training a single boundary predictor and applying it to all images regardless of their content is bound to be suboptimal. In this paper we therefore propose situational object boundary detection: We first define a variety of situations and train a specialized object boundary detector for each of them using [Dollar and Zitnick 2013]. Then given a test image, we classify it into these situations using its context, which we model by global image appearance. We apply the corresponding situational object boundary detectors, and fuse them based on the classification probabilities. In experiments on ImageNet, Microsoft COCO, and Pascal VOC 2012 segmentation we show that our situational object boundary detection gives significant improvements over a monolithic approach. Additionally, our method substantially outperforms [Hariharan et al. 2011] on semantic contour detection on their SBD dataset.