Thomas Mensink

Apprentissage de modèles pour la classification et la recherche d'images: Nous assistons actuellement à une explosion de la quantité des données visuelles. Par exemple, plusieurs millions de photos sont partagées quotidiennement sur les réseaux sociaux. Les méthodes d'interprétation d'images vise à faciliter l'accès à ces données visuelles, d'une manière sémantiquement compréhensible. Dans ce manuscrit, nous définissons certains buts détaillés qui sont intéressants pour les taches d'interprétation d'images, telles que la classification ou la recherche d'images, que nous considérons dans les trois chapitres principaux. Tout d'abord, nous visons l'exploitation de la nature multimodale de nombreuses bases de données, pour lesquelles les documents sont composés d'images et de descriptions textuelles. Dans ce but, nous définissons des similarités entre le contenu visuel d'un document, et la description textuelle d'un autre document. Ces similarités sont calculées en deux étapes, tout d'abord nous trouvons les voisins visuellement similaires dans la base multimodale, puis nous utilisons les descriptions textuelles de ces voisins afin de définir une similarité avec la description textuelle de n'importe quel document. Ensuite, nous présentons une série de modèles structurés pour la classification d'images, qui encodent explicitement les interactions binaires entre les étiquettes (ou labels). Ces modèles sont plus expressifs que des prédicateurs d'étiquette indépendants, et aboutissent à des prédictions plus fiables, en particulier dans un scenario de prédiction interactive, où les utilisateurs fournissent les valeurs de certaines des étiquettes d'images. Un scenario interactif comme celui-ci offre un compromis intéressant entre la précision, et l'effort d'annotation manuelle requis. Nous explorons les modèles structurés pour la classification multi-étiquette d'images, pour la classification d'image basée sur les attributs, et pour l'optimisation de certaines mesures de rang spécifiques. Enfin, nous explorons les classifieurs par k plus proches voisins, et les classifieurs par plus proche moyenne, pour la classification d'images à grande échelle. Nous proposons des méthodes d'apprentissage de métrique efficaces pour améliorer les performances de classification, et appliquons ces méthodes à une base de plus d'un million d'images d'apprentissage, et d'un millier de classes. Comme les deux méthodes de classification permettent d'incorporer des classes non vues pendant l'apprentissage à un coût presque nul, nous avons également étudié leur performance pour la généralisation. Nous montrons que la classification par plus proche moyenne généralise à partir d'un millier de classes, sur dix mille classes à un coût négligeable, et les performances obtenus sont comparables à l'état de l'art.

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.

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 .

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.

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.

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.

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 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.

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.

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.