Diane Larlus

Vadim Tschernezki, Iro Laina, Diane Larlus et al.

We present Neural Feature Fusion Fields (N3F), a method that improves dense 2D image feature extractors when the latter are applied to the analysis of multiple images reconstructible as a 3D scene. Given an image feature extractor, for example pre-trained using self-supervision, N3F uses it as a teacher to learn a student network defined in 3D space. The 3D student network is similar to a neural radiance field that distills said features and can be trained with the usual differentiable rendering machinery. As a consequence, N3F is readily applicable to most neural rendering formulations, including vanilla NeRF and its extensions to complex dynamic scenes. We show that our method not only enables semantic understanding in the context of scene-specific neural fields without the use of manual labels, but also consistently improves over the self-supervised 2D baselines. This is demonstrated by considering various tasks, such as 2D object retrieval, 3D segmentation, and scene editing, in diverse sequences, including long egocentric videos in the EPIC-KITCHENS benchmark.

Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu et al.

Neural rendering is fuelling a unification of learning, 3D geometry and video understanding that has been waiting for more than two decades. Progress, however, is still hampered by a lack of suitable datasets and benchmarks. To address this gap, we introduce EPIC Fields, an augmentation of EPIC-KITCHENS with 3D camera information. Like other datasets for neural rendering, EPIC Fields removes the complex and expensive step of reconstructing cameras using photogrammetry, and allows researchers to focus on modelling problems. We illustrate the challenge of photogrammetry in egocentric videos of dynamic actions and propose innovations to address them. Compared to other neural rendering datasets, EPIC Fields is better tailored to video understanding because it is paired with labelled action segments and the recent VISOR segment annotations. To further motivate the community, we also evaluate two benchmark tasks in neural rendering and segmenting dynamic objects, with strong baselines that showcase what is not possible today. We also highlight the advantage of geometry in semi-supervised video object segmentations on the VISOR annotations. EPIC Fields reconstructs 96% of videos in EPICKITCHENS, registering 19M frames in 99 hours recorded in 45 kitchens.

Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus et al.

Recent image generation models such as Stable Diffusion have exhibited an impressive ability to generate fairly realistic images starting from a simple text prompt. Could such models render real images obsolete for training image prediction models? In this paper, we answer part of this provocative question by investigating the need for real images when training models for ImageNet classification. Provided only with the class names that have been used to build the dataset, we explore the ability of Stable Diffusion to generate synthetic clones of ImageNet and measure how useful these are for training classification models from scratch. We show that with minimal and class-agnostic prompt engineering, ImageNet clones are able to close a large part of the gap between models produced by synthetic images and models trained with real images, for the several standard classification benchmarks that we consider in this study. More importantly, we show that models trained on synthetic images exhibit strong generalization properties and perform on par with models trained on real data for transfer. Project page: https://europe.naverlabs.com/imagenet-sd/

Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka et al.

An intuitive way to search for images is to use queries composed of an example image and a complementary text. While the first provides rich and implicit context for the search, the latter explicitly calls for new traits, or specifies how some elements of the example image should be changed to retrieve the desired target image. Current approaches typically combine the features of each of the two elements of the query into a single representation, which can then be compared to the ones of the potential target images. Our work aims at shedding new light on the task by looking at it through the prism of two familiar and related frameworks: text-to-image and image-to-image retrieval. Taking inspiration from them, we exploit the specific relation of each query element with the targeted image and derive light-weight attention mechanisms which enable to mediate between the two complementary modalities. We validate our approach on several retrieval benchmarks, querying with images and their associated free-form text modifiers. Our method obtains state-of-the-art results without resorting to side information, multi-level features, heavy pre-training nor large architectures as in previous works.

Mert Bulent Sariyildiz, Yannis Kalantidis, Karteek Alahari et al.

We consider the problem of training a deep neural network on a given classification task, e.g., ImageNet-1K (IN1K), so that it excels at both the training task as well as at other (future) transfer tasks. These two seemingly contradictory properties impose a trade-off between improving the model's generalization and maintaining its performance on the original task. Models trained with self-supervised learning tend to generalize better than their supervised counterparts for transfer learning; yet, they still lag behind supervised models on IN1K. In this paper, we propose a supervised learning setup that leverages the best of both worlds. We extensively analyze supervised training using multi-scale crops for data augmentation and an expendable projector head, and reveal that the design of the projector allows us to control the trade-off between performance on the training task and transferability. We further replace the last layer of class weights with class prototypes computed on the fly using a memory bank and derive two models: t-ReX that achieves a new state of the art for transfer learning and outperforms top methods such as DINO and PAWS on IN1K, and t-ReX* that matches the highly optimized RSB-A1 model on IN1K while performing better on transfer tasks. Code and pretrained models: https://europe.naverlabs.com/t-rex

Juliette Marrie, Michael Arbel, Diane Larlus et al.

Data augmentation is known to improve the generalization capabilities of neural networks, provided that the set of transformations is chosen with care, a selection often performed manually. Automatic data augmentation aims at automating this process. However, most recent approaches still rely on some prior information; they start from a small pool of manually-selected default transformations that are either used to pretrain the network or forced to be part of the policy learned by the automatic data augmentation algorithm. In this paper, we propose to directly learn the augmentation policy without leveraging such prior knowledge. The resulting bilevel optimization problem becomes more challenging due to the larger search space and the inherent instability of bilevel optimization algorithms. To mitigate these issues (i) we follow a successive cold-start strategy with a Kullback-Leibler regularization, and (ii) we parameterize magnitudes as continuous distributions. Our approach leads to competitive results on standard benchmarks despite a more challenging setting, and generalizes beyond natural images.

Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas et al.

Pretrained models have become a commodity and offer strong results on a broad range of tasks. In this work, we focus on classification and seek to learn a unique encoder able to take from several complementary pretrained models. We aim at even stronger generalization across a variety of classification tasks. We propose to learn such an encoder via multi-teacher distillation. We first thoroughly analyse standard distillation when driven by multiple strong teachers with complementary strengths. Guided by this analysis, we gradually propose improvements to the basic distillation setup. Among those, we enrich the architecture of the encoder with a ladder of expendable projectors, which increases the impact of intermediate features during distillation, and we introduce teacher dropping, a regularization mechanism that better balances the teachers' influence. Our final distillation strategy leads to student models of the same capacity as any of the teachers, while retaining or improving upon the performance of the best teacher for each task. Project page and code: https://europe.naverlabs.com/unic

Jon Almazán, Byungsoo Ko, Geonmo Gu et al.

Strong image search models can be learned for a specific domain, ie. set of labels, provided that some labeled images of that domain are available. A practical visual search model, however, should be versatile enough to solve multiple retrieval tasks simultaneously, even if those cover very different specialized domains. Additionally, it should be able to benefit from even unlabeled images from these various retrieval tasks. This is the more practical scenario that we consider in this paper. We address it with the proposed Grappa, an approach that starts from a strong pretrained model, and adapts it to tackle multiple retrieval tasks concurrently, using only unlabeled images from the different task domains. We extend the pretrained model with multiple independently trained sets of adaptors that use pseudo-label sets of different sizes, effectively mimicking different pseudo-granularities. We reconcile all adaptor sets into a single unified model suited for all retrieval tasks by learning fusion layers that we guide by propagating pseudo-granularity attentions across neighbors in the feature space. Results on a benchmark composed of six heterogeneous retrieval tasks show that the unsupervised Grappa model improves the zero-shot performance of a state-of-the-art self-supervised learning model, and in some places reaches or improves over a task label-aware oracle that selects the most fitting pseudo-granularity per task.

Gabriela Csurka, Tyler L. Hayes, Diane Larlus et al.

Deep learning models are effective, yet brittle. Even carefully trained, their behavior tends to be hard to predict when confronted with out-of-distribution samples. In this work, our goal is to propose a simple yet effective solution to predict and describe via natural language potential failure modes of computer vision models. Given a pretrained model and a set of samples, our aim is to find sentences that accurately describe the visual conditions in which the model underperforms. In order to study this important topic and foster future research on it, we formalize the problem of Language-Based Error Explainability (LBEE) and propose a set of metrics to evaluate and compare different methods for this task. We propose solutions that operate in a joint vision-and-language embedding space, and can characterize through language descriptions model failures caused, e.g., by objects unseen during training or adverse visual conditions. We experiment with different tasks, such as classification under the presence of dataset bias and semantic segmentation in unseen environments, and show that the proposed methodology isolates nontrivial sentences associated with specific error causes. We hope our work will help practitioners better understand the behavior of models, increasing their overall safety and interpretability.

Pau de Jorge, César Roberto de Souza, Björn Michele et al.

Efficiently merging several models fine-tuned for different tasks, but stemming from the same pretrained base model, is of great practical interest. Despite extensive prior work, most evaluations of model merging in computer vision are restricted to image classification using CLIP, where different classification datasets define different tasks. In this work, our goal is to make model merging more practical and show its relevance on challenging scenarios beyond this specific setting. In most vision scenarios, different tasks rely on trainable and usually heterogeneous decoders. Differently from previous studies with frozen decoders, where merged models can be evaluated right away, the non-trivial cost of decoder training renders hyperparameter selection based on downstream performance impractical. To address this, we introduce the task alignment proxy, and show how it can be used to speed up hyperparameter selection by orders of magnitude while retaining performance. Equipped with the task alignment proxy, we extend the applicability of model merging to multi-task vision models beyond CLIP-based classification.

Zeren Jiang, Chuanxia Zheng, Iro Laina et al.

We propose Mesh4D, a feed-forward model for monocular 4D mesh reconstruction. Given a monocular video of a dynamic object, our model reconstructs the object's complete 3D shape and motion, represented as a deformation field. Our key contribution is a compact latent space that encodes the entire animation sequence in a single pass. This latent space is learned by an autoencoder that, during training, is guided by the skeletal structure of the training objects, providing strong priors on plausible deformations. Crucially, skeletal information is not required at inference time. The encoder employs spatio-temporal attention, yielding a more stable representation of the object's overall deformation. Building on this representation, we train a latent diffusion model that, conditioned on the input video and the mesh reconstructed from the first frame, predicts the full animation in one shot. We evaluate Mesh4D on reconstruction and novel view synthesis benchmarks, outperforming prior methods in recovering accurate 3D shape and deformation.

Philippe Weinzaepfel, Thomas Lucas, Diane Larlus et al.

Methods that combine local and global features have recently shown excellent performance on multiple challenging deep image retrieval benchmarks, but their use of local features raises at least two issues. First, these local features simply boil down to the localized map activations of a neural network, and hence can be extremely redundant. Second, they are typically trained with a global loss that only acts on top of an aggregation of local features; by contrast, testing is based on local feature matching, which creates a discrepancy between training and testing. In this paper, we propose a novel architecture for deep image retrieval, based solely on mid-level features that we call Super-features. These Super-features are constructed by an iterative attention module and constitute an ordered set in which each element focuses on a localized and discriminant image pattern. For training, they require only image labels. A contrastive loss operates directly at the level of Super-features and focuses on those that match across images. A second complementary loss encourages diversity. Experiments on common landmark retrieval benchmarks validate that Super-features substantially outperform state-of-the-art methods when using the same number of features, and only require a significantly smaller memory footprint to match their performance. Code and models are available at: https://github.com/naver/FIRe.

Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio de Rezende et al.

Cross-modal retrieval methods build a common representation space for samples from multiple modalities, typically from the vision and the language domains. For images and their captions, the multiplicity of the correspondences makes the task particularly challenging. Given an image (respectively a caption), there are multiple captions (respectively images) that equally make sense. In this paper, we argue that deterministic functions are not sufficiently powerful to capture such one-to-many correspondences. Instead, we propose to use Probabilistic Cross-Modal Embedding (PCME), where samples from the different modalities are represented as probabilistic distributions in the common embedding space. Since common benchmarks such as COCO suffer from non-exhaustive annotations for cross-modal matches, we propose to additionally evaluate retrieval on the CUB dataset, a smaller yet clean database where all possible image-caption pairs are annotated. We extensively ablate PCME and demonstrate that it not only improves the retrieval performance over its deterministic counterpart but also provides uncertainty estimates that render the embeddings more interpretable. Code is available at https://github.com/naver-ai/pcme

Zeren Jiang, Yushi Lan, Yihang Luo et al.

Dense 3D reconstruction and tracking of dynamic scenes from monocular video remains an important open challenge in computer vision. Progress in this area has been constrained by the scarcity of high-quality datasets with dense, complete, and accurate geometric annotations. To address this limitation, we introduce Syn4D, a multiview synthetic dataset of dynamic scenes that includes ground-truth camera motion, depth maps, dense tracking, and parametric human pose annotations. A key feature of Syn4D is the ability to unproject any pixel into 3D to any time and to any camera. We conduct extensive evaluations across multiple downstream tasks to demonstrate the utility and effectiveness of the proposed dataset, including 4D scene reconstruction, 3D point tracking, geometry-aware camera retargeting, and human pose estimation. The experimental results highlight Syn4D's potential to facilitate research in dynamic scene understanding and spatiotemporal modeling.

Zeren Jiang, Chuanxia Zheng, Iro Laina et al.

We introduce Geo4D, a method to repurpose video diffusion models for monocular 3D reconstruction of dynamic scenes. By leveraging the strong dynamic priors captured by large-scale pre-trained video models, Geo4D can be trained using only synthetic data while generalizing well to real data in a zero-shot manner. Geo4D predicts several complementary geometric modalities, namely point, disparity, and ray maps. We propose a new multi-modal alignment algorithm to align and fuse these modalities, as well as a sliding window approach at inference time, thus enabling robust and accurate 4D reconstruction of long videos. Extensive experiments across multiple benchmarks show that Geo4D significantly surpasses state-of-the-art video depth estimation methods.

Juliette Marrie, Romain Menegaux, Michael Arbel et al.

We address the problem of extending the capabilities of vision foundation models such as DINO, SAM, and CLIP, to 3D tasks. Specifically, we introduce a novel method to uplift 2D image features into Gaussian Splatting representations of 3D scenes. Unlike traditional approaches that rely on minimizing a reconstruction loss, our method employs a simpler and more efficient feature aggregation technique, augmented by a graph diffusion mechanism. Graph diffusion refines 3D features, such as coarse segmentation masks, by leveraging 3D geometry and pairwise similarities induced by DINOv2. Our approach achieves performance comparable to the state of the art on multiple downstream tasks while delivering significant speed-ups. Notably, we obtain competitive segmentation results using only generic DINOv2 features, despite DINOv2 not being trained on millions of annotated segmentation masks like SAM. When applied to CLIP features, our method demonstrates strong performance in open-vocabulary object segmentation tasks, highlighting the versatility of our approach.

Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas et al.

Recent multi-teacher distillation methods have unified the encoders of multiple foundation models into a single encoder, achieving competitive performance on core vision tasks like classification, segmentation, and depth estimation. This led us to ask: Could similar success be achieved when the pool of teachers also includes vision models specialized in diverse tasks across both 2D and 3D perception? In this paper, we define and investigate the problem of heterogeneous teacher distillation, or co-distillation, a challenging multi-teacher distillation scenario where teacher models vary significantly in both (a) their design objectives and (b) the data they were trained on. We explore data-sharing strategies and teacher-specific encoding, and introduce DUNE, a single encoder excelling in 2D vision, 3D understanding, and 3D human perception. Our model achieves performance comparable to that of its larger teachers, sometimes even outperforming them, on their respective tasks. Notably, DUNE surpasses MASt3R in Map-free Visual Relocalization with a much smaller encoder.

Tyler L. Hayes, César R. de Souza, Namil Kim et al.

Object detectors are typically trained once and for all on a fixed set of classes. However, this closed-world assumption is unrealistic in practice, as new classes will inevitably emerge after the detector is deployed in the wild. In this work, we look at ways to extend a detector trained for a set of base classes so it can i) spot the presence of novel classes, and ii) automatically enrich its repertoire to be able to detect those newly discovered classes together with the base ones. We propose PANDAS, a method for novel class discovery and detection. It discovers clusters representing novel classes from unlabeled data, and represents old and new classes with prototypes. During inference, a distance-based classifier uses these prototypes to assign a label to each detected object instance. The simplicity of our method makes it widely applicable. We experimentally demonstrate the effectiveness of PANDAS on the VOC 2012 and COCO-to-LVIS benchmarks. It performs favorably against the state of the art for this task while being computationally more affordable.

Juliette Marrie, Michael Arbel, Julien Mairal et al.

Large pretrained visual models exhibit remarkable generalization across diverse recognition tasks. Yet, real-world applications often demand compact models tailored to specific problems. Variants of knowledge distillation have been devised for such a purpose, enabling task-specific compact models (the students) to learn from a generic large pretrained one (the teacher). In this paper, we show that the excellent robustness and versatility of recent pretrained models challenge common practices established in the literature, calling for a new set of optimal guidelines for task-specific distillation. To address the lack of samples in downstream tasks, we also show that a variant of Mixup based on stable diffusion complements standard data augmentation. This strategy eliminates the need for engineered text prompts and improves distillation of generic models into streamlined specialized networks.

Vadim Tschernezki, Diane Larlus, Iro Laina et al.

Computer vision is largely based on 2D techniques, with 3D vision still relegated to a relatively narrow subset of applications. However, by building on recent advances in 3D models such as neural radiance fields, some authors have shown that 3D techniques can at last improve outputs extracted from independent 2D views, by fusing them into 3D and denoising them. This is particularly helpful in egocentric videos, where the camera motion is significant, but only under the assumption that the scene itself is static. In fact, as shown in the recent analysis conducted by EPIC Fields, 3D techniques are ineffective when it comes to studying dynamic phenomena, and, in particular, when segmenting moving objects. In this paper, we look into this issue in more detail. First, we propose to improve dynamic segmentation in 3D by fusing motion segmentation predictions from a 2D-based model into layered radiance fields (Layered Motion Fusion). However, the high complexity of long, dynamic videos makes it challenging to capture the underlying geometric structure, and, as a result, hinders the fusion of motion cues into the (incomplete) scene geometry. We address this issue through test-time refinement, which helps the model to focus on specific frames, thereby reducing the data complexity. This results in a synergy between motion fusion and the refinement, and in turn leads to segmentation predictions of the 3D model that surpass the 2D baseline by a large margin. This demonstrates that 3D techniques can enhance 2D analysis even for dynamic phenomena in a challenging and realistic setting.

Yannis Kalantidis, Mert Bülent Sarıyıldız, Rafael S. Rezende et al.

State-of-the-art visual localization approaches generally rely on a first image retrieval step whose role is crucial. Yet, retrieval often struggles when facing varying conditions, due to e.g. weather or time of day, with dramatic consequences on the visual localization accuracy. In this paper, we improve this retrieval step and tailor it to the final localization task. Among the several changes we advocate for, we propose to synthesize variants of the training set images, obtained from generative text-to-image models, in order to automatically expand the training set towards a number of nameable variations that particularly hurt visual localization. After expanding the training set, we propose a training approach that leverages the specificities and the underlying geometry of this mix of real and synthetic images. We experimentally show that those changes translate into large improvements for the most challenging visual localization datasets. Project page: https://europe.naverlabs.com/ret4loc

Subhankar Roy, Riccardo Volpi, Gabriela Csurka et al.

Class-incremental semantic image segmentation assumes multiple model updates, each enriching the model to segment new categories. This is typically carried out by providing expensive pixel-level annotations to the training algorithm for all new objects, limiting the adoption of such methods in practical applications. Approaches that solely require image-level labels offer an attractive alternative, yet, such coarse annotations lack precise information about the location and boundary of the new objects. In this paper we argue that, since classes represent not just indices but semantic entities, the conceptual relationships between them can provide valuable information that should be leveraged. We propose a weakly supervised approach that exploits such semantic relations to transfer objectness prior from the previously learned classes into the new ones, complementing the supervisory signal from image-level labels. We validate our approach on a number of continual learning tasks, and show how even a simple pairwise interaction between classes can significantly improve the segmentation mask quality of both old and new classes. We show these conclusions still hold for longer and, hence, more realistic sequences of tasks and for a challenging few-shot scenario.

Riccardo Volpi, Pau de Jorge, Diane Larlus et al.

We propose a new problem formulation and a corresponding evaluation framework to advance research on unsupervised domain adaptation for semantic image segmentation. The overall goal is fostering the development of adaptive learning systems that will continuously learn, without supervision, in ever-changing environments. Typical protocols that study adaptation algorithms for segmentation models are limited to few domains, adaptation happens offline, and human intervention is generally required, at least to annotate data for hyper-parameter tuning. We argue that such constraints are incompatible with algorithms that can continuously adapt to different real-world situations. To address this, we propose a protocol where models need to learn online, from sequences of temporally correlated images, requiring continuous, frame-by-frame adaptation. We accompany this new protocol with a variety of baselines to tackle the proposed formulation, as well as an extensive analysis of their behaviors, which can serve as a starting point for future research.

Jonathan Munro, Michael Wray, Diane Larlus et al.

Given a gallery of uncaptioned video sequences, this paper considers the task of retrieving videos based on their relevance to an unseen text query. To compensate for the lack of annotations, we rely instead on a related video gallery composed of video-caption pairs, termed the source gallery, albeit with a domain gap between its videos and those in the target gallery. We thus introduce the problem of Unsupervised Domain Adaptation for Cross-modal Video Retrieval, along with a new benchmark on fine-grained actions. We propose a novel iterative domain alignment method by means of pseudo-labelling target videos and cross-domain (i.e. source-target) ranking. Our approach adapts the embedding space to the target gallery, consistently outperforming source-only as well as marginal and conditional alignment methods.

Vadim Tschernezki, Diane Larlus, Andrea Vedaldi

Given a raw video sequence taken from a freely-moving camera, we study the problem of decomposing the observed 3D scene into a static background and a dynamic foreground containing the objects that move in the video sequence. This task is reminiscent of the classic background subtraction problem, but is significantly harder because all parts of the scene, static and dynamic, generate a large apparent motion due to the camera large viewpoint change. In particular, we consider egocentric videos and further separate the dynamic component into objects and the actor that observes and moves them. We achieve this factorization by reconstructing the video via a triple-stream neural rendering network that explains the different motions based on corresponding inductive biases. We demonstrate that our method can successfully separate the different types of motion, outperforming recent neural rendering baselines at this task, and can accurately segment moving objects. We do so by assessing the method empirically on challenging videos from the EPIC-KITCHENS dataset which we augment with appropriate annotations to create a new benchmark for the task of dynamic object segmentation on unconstrained video sequences, for complex 3D environments.

Yannis Kalantidis, Carlos Lassance, Jon Almazan et al.

Dimensionality reduction methods are unsupervised approaches which learn low-dimensional spaces where some properties of the initial space, typically the notion of "neighborhood", are preserved. Such methods usually require propagation on large k-NN graphs or complicated optimization solvers. On the other hand, self-supervised learning approaches, typically used to learn representations from scratch, rely on simple and more scalable frameworks for learning. In this paper, we propose TLDR, a dimensionality reduction method for generic input spaces that is porting the recent self-supervised learning framework of Zbontar et al. (2021) to the specific task of dimensionality reduction, over arbitrary representations. We propose to use nearest neighbors to build pairs from a training set and a redundancy reduction loss to learn an encoder that produces representations invariant across such pairs. TLDR is a method that is simple, easy to train, and of broad applicability; it consists of an offline nearest neighbor computation step that can be highly approximated, and a straightforward learning process. Aiming for scalability, we focus on improving linear dimensionality reduction, and show consistent gains on image and document retrieval tasks, e.g. gaining +4% mAP over PCA on ROxford for GeM- AP, improving the performance of DINO on ImageNet or retaining it with a 10x compression.

Mert Bulent Sariyildiz, Yannis Kalantidis, Diane Larlus et al.

Measuring concept generalization, i.e., the extent to which models trained on a set of (seen) visual concepts can be leveraged to recognize a new set of (unseen) concepts, is a popular way of evaluating visual representations, especially in a self-supervised learning framework. Nonetheless, the choice of unseen concepts for such an evaluation is usually made arbitrarily, and independently from the seen concepts used to train representations, thus ignoring any semantic relationships between the two. In this paper, we argue that the semantic relationships between seen and unseen concepts affect generalization performance and propose ImageNet-CoG, a novel benchmark on the ImageNet-21K (IN-21K) dataset that enables measuring concept generalization in a principled way. Our benchmark leverages expert knowledge that comes from WordNet in order to define a sequence of unseen IN-21K concept sets that are semantically more and more distant from the ImageNet-1K (IN-1K) subset, a ubiquitous training set. This allows us to benchmark visual representations learned on IN-1K out-of-the box. We conduct a large-scale study encompassing 31 convolution and transformer-based models and show how different architectures, levels of supervision, regularization techniques and use of web data impact the concept generalization performance.

Andrés Mafla, Rafael Sampaio de Rezende, Lluís Gómez et al.

Recent models for cross-modal retrieval have benefited from an increasingly rich understanding of visual scenes, afforded by scene graphs and object interactions to mention a few. This has resulted in an improved matching between the visual representation of an image and the textual representation of its caption. Yet, current visual representations overlook a key aspect: the text appearing in images, which may contain crucial information for retrieval. In this paper, we first propose a new dataset that allows exploration of cross-modal retrieval where images contain scene-text instances. Then, armed with this dataset, we describe several approaches which leverage scene text, including a better scene-text aware cross-modal retrieval method which uses specialized representations for text from the captions and text from the visual scene, and reconcile them in a common embedding space. Extensive experiments confirm that cross-modal retrieval approaches benefit from scene text and highlight interesting research questions worth exploring further. Dataset and code are available at http://europe.naverlabs.com/stacmr

Riccardo Volpi, Diane Larlus, Grégory Rogez

Most standard learning approaches lead to fragile models which are prone to drift when sequentially trained on samples of a different nature - the well-known "catastrophic forgetting" issue. In particular, when a model consecutively learns from different visual domains, it tends to forget the past domains in favor of the most recent ones. In this context, we show that one way to learn models that are inherently more robust against forgetting is domain randomization - for vision tasks, randomizing the current domain's distribution with heavy image manipulations. Building on this result, we devise a meta-learning strategy where a regularizer explicitly penalizes any loss associated with transferring the model from the current domain to different "auxiliary" meta-domains, while also easing adaptation to them. Such meta-domains are also generated through randomized image manipulations. We empirically demonstrate in a variety of experiments - spanning from classification to semantic segmentation - that our approach results in models that are less prone to catastrophic forgetting when transferred to new domains.

Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion et al.

Contrastive learning has become a key component of self-supervised learning approaches for computer vision. By learning to embed two augmented versions of the same image close to each other and to push the embeddings of different images apart, one can train highly transferable visual representations. As revealed by recent studies, heavy data augmentation and large sets of negatives are both crucial in learning such representations. At the same time, data mixing strategies either at the image or the feature level improve both supervised and semi-supervised learning by synthesizing novel examples, forcing networks to learn more robust features. In this paper, we argue that an important aspect of contrastive learning, i.e., the effect of hard negatives, has so far been neglected. To get more meaningful negative samples, current top contrastive self-supervised learning approaches either substantially increase the batch sizes, or keep very large memory banks; increasing the memory size, however, leads to diminishing returns in terms of performance. We therefore start by delving deeper into a top-performing framework and show evidence that harder negatives are needed to facilitate better and faster learning. Based on these observations, and motivated by the success of data mixing, we propose hard negative mixing strategies at the feature level, that can be computed on-the-fly with a minimal computational overhead. We exhaustively ablate our approach on linear classification, object detection and instance segmentation and show that employing our hard negative mixing procedure improves the quality of visual representations learned by a state-of-the-art self-supervised learning method.

Mert Bulent Sariyildiz, Julien Perez, Diane Larlus

Pretraining general-purpose visual features has become a crucial part of tackling many computer vision tasks. While one can learn such features on the extensively-annotated ImageNet dataset, recent approaches have looked at ways to allow for noisy, fewer, or even no annotations to perform such pretraining. Starting from the observation that captioned images are easily crawlable, we argue that this overlooked source of information can be exploited to supervise the training of visual representations. To do so, motivated by the recent progresses in language models, we introduce {\em image-conditioned masked language modeling} (ICMLM) -- a proxy task to learn visual representations over image-caption pairs. ICMLM consists in predicting masked words in captions by relying on visual cues. To tackle this task, we propose hybrid models, with dedicated visual and textual encoders, and we show that the visual representations learned as a by-product of solving this task transfer well to a variety of target tasks. Our experiments confirm that image captions can be leveraged to inject global and localized semantic information into visual representations. Project website: https://europe.naverlabs.com/icmlm.

Michael Wray, Diane Larlus, Gabriela Csurka et al.

We address the problem of cross-modal fine-grained action retrieval between text and video. Cross-modal retrieval is commonly achieved through learning a shared embedding space, that can indifferently embed modalities. In this paper, we propose to enrich the embedding by disentangling parts-of-speech (PoS) in the accompanying captions. We build a separate multi-modal embedding space for each PoS tag. The outputs of multiple PoS embeddings are then used as input to an integrated multi-modal space, where we perform action retrieval. All embeddings are trained jointly through a combination of PoS-aware and PoS-agnostic losses. Our proposal enables learning specialised embedding spaces that offer multiple views of the same embedded entities. We report the first retrieval results on fine-grained actions for the large-scale EPIC dataset, in a generalised zero-shot setting. Results show the advantage of our approach for both video-to-text and text-to-video action retrieval. We also demonstrate the benefit of disentangling the PoS for the generic task of cross-modal video retrieval on the MSR-VTT dataset.

David Novotny, Samuel Albanie, Diane Larlus et al.

Object detection and instance segmentation are dominated by region-based methods such as Mask RCNN. However, there is a growing interest in reducing these problems to pixel labeling tasks, as the latter could be more efficient, could be integrated seamlessly in image-to-image network architectures as used in many other tasks, and could be more accurate for objects that are not well approximated by bounding boxes. In this paper we show theoretically and empirically that constructing dense pixel embeddings that can separate object instances cannot be easily achieved using convolutional operators. At the same time, we show that simple modifications, which we call semi-convolutional, have a much better chance of succeeding at this task. We use the latter to show a connection to Hough voting as well as to a variant of the bilateral kernel that is spatially steered by a convolutional network. We demonstrate that these operators can also be used to improve approaches such as Mask RCNN, demonstrating better segmentation of complex biological shapes and PASCAL VOC categories than achievable by Mask RCNN alone.

David Novotny, Samuel Albanie, Diane Larlus et al.

Self-supervision can dramatically cut back the amount of manually-labelled data required to train deep neural networks. While self-supervision has usually been considered for tasks such as image classification, in this paper we aim at extending it to geometry-oriented tasks such as semantic matching and part detection. We do so by building on several recent ideas in unsupervised landmark detection. Our approach learns dense distinctive visual descriptors from an unlabelled dataset of images using synthetic image transformations. It does so by means of a robust probabilistic formulation that can introspectively determine which image regions are likely to result in stable image matching. We show empirically that a network pre-trained in this manner requires significantly less supervision to learn semantic object parts compared to numerous pre-training alternatives. We also show that the pre-trained representation is excellent for semantic object matching.

Jon Almazan, Bojana Gajic, Naila Murray et al.

Training a deep architecture using a ranking loss has become standard for the person re-identification task. Increasingly, these deep architectures include additional components that leverage part detections, attribute predictions, pose estimators and other auxiliary information, in order to more effectively localize and align discriminative image regions. In this paper we adopt a different approach and carefully design each component of a simple deep architecture and, critically, the strategy for training it effectively for person re-identification. We extensively evaluate each design choice, leading to a list of good practices for person re-identification. By following these practices, our approach outperforms the state of the art, including more complex methods with auxiliary components, by large margins on four benchmark datasets. We also provide a qualitative analysis of our trained representation which indicates that, while compact, it is able to capture information from localized and discriminative regions, in a manner akin to an implicit attention mechanism.

David Novotny, Diane Larlus, Andrea Vedaldi

Traditional approaches for learning 3D object categories use either synthetic data or manual supervision. In this paper, we propose a method which does not require manual annotations and is instead cued by observing objects from a moving vantage point. Our system builds on two innovations: a Siamese viewpoint factorization network that robustly aligns different videos together without explicitly comparing 3D shapes; and a 3D shape completion network that can extract the full shape of an object from partial observations. We also demonstrate the benefits of configuring networks to perform probabilistic predictions as well as of geometry-aware data augmentation schemes. We obtain state-of-the-art results on publicly-available benchmarks.

David Novotny, Diane Larlus, Andrea Vedaldi

Despite significant progress of deep learning in recent years, state-of-the-art semantic matching methods still rely on legacy features such as SIFT or HoG. We argue that the strong invariance properties that are key to the success of recent deep architectures on the classification task make them unfit for dense correspondence tasks, unless a large amount of supervision is used. In this work, we propose a deep network, termed AnchorNet, that produces image representations that are well-suited for semantic matching. It relies on a set of filters whose response is geometrically consistent across different object instances, even in the presence of strong intra-class, scale, or viewpoint variations. Trained only with weak image-level labels, the final representation successfully captures information about the object structure and improves results of state-of-the-art semantic matching methods such as the deformable spatial pyramid or the proposal flow methods. We show positive results on the cross-instance matching task where different instances of the same object category are matched as well as on a new cross-category semantic matching task aligning pairs of instances each from a different object class.

Albert Gordo, Jon Almazan, Jerome Revaud et al.

While deep learning has become a key ingredient in the top performing methods for many computer vision tasks, it has failed so far to bring similar improvements to instance-level image retrieval. In this article, we argue that reasons for the underwhelming results of deep methods on image retrieval are threefold: i) noisy training data, ii) inappropriate deep architecture, and iii) suboptimal training procedure. We address all three issues. First, we leverage a large-scale but noisy landmark dataset and develop an automatic cleaning method that produces a suitable training set for deep retrieval. Second, we build on the recent R-MAC descriptor, show that it can be interpreted as a deep and differentiable architecture, and present improvements to enhance it. Last, we train this network with a siamese architecture that combines three streams with a triplet loss. At the end of the training process, the proposed architecture produces a global image representation in a single forward pass that is well suited for image retrieval. Extensive experiments show that our approach significantly outperforms previous retrieval approaches, including state-of-the-art methods based on costly local descriptor indexing and spatial verification. On Oxford 5k, Paris 6k and Holidays, we respectively report 94.7, 96.6, and 94.8 mean average precision. Our representations can also be heavily compressed using product quantization with little loss in accuracy. For additional material, please see www.xrce.xerox.com/Deep-Image-Retrieval.

David Novotny, Diane Larlus, Andrea Vedaldi

While recent research in image understanding has often focused on recognizing more types of objects, understanding more about the objects is just as important. Recognizing object parts and attributes has been extensively studied before, yet learning large space of such concepts remains elusive due to the high cost of providing detailed object annotations for supervision. The key contribution of this paper is an algorithm to learn the nameable parts of objects automatically, from images obtained by querying Web search engines. The key challenge is the high level of noise in the annotations; to address it, we propose a new unified embedding space where the appearance and geometry of objects and their semantic parts are represented uniformly. Geometric relationships are induced in a soft manner by a rich set of nonsemantic mid-level anchors, bridging the gap between semantic and non-semantic parts. We also show that the resulting embedding provides a visually-intuitive mechanism to navigate the learned concepts and their corresponding images.

Albert Gordo, Jon Almazan, Jerome Revaud et al.

We propose a novel approach for instance-level image retrieval. It produces a global and compact fixed-length representation for each image by aggregating many region-wise descriptors. In contrast to previous works employing pre-trained deep networks as a black box to produce features, our method leverages a deep architecture trained for the specific task of image retrieval. Our contribution is twofold: (i) we leverage a ranking framework to learn convolution and projection weights that are used to build the region features; and (ii) we employ a region proposal network to learn which regions should be pooled to form the final global descriptor. We show that using clean training data is key to the success of our approach. To that aim, we use a large scale but noisy landmark dataset and develop an automatic cleaning approach. The proposed architecture produces a global image representation in a single forward pass. Our approach significantly outperforms previous approaches based on global descriptors on standard datasets. It even surpasses most prior works based on costly local descriptor indexing and spatial verification. Additional material is available at www.xrce.xerox.com/Deep-Image-Retrieval.

Gabriela Csurka, Diane Larlus, Albert Gordo et al.

In this article we study the problem of document image representation based on visual features. We propose a comprehensive experimental study that compares three types of visual document image representations: (1) traditional so-called shallow features, such as the RunLength and the Fisher-Vector descriptors, (2) deep features based on Convolutional Neural Networks, and (3) features extracted from hybrid architectures that take inspiration from the two previous ones. We evaluate these features in several tasks (i.e. classification, clustering, and retrieval) and in different setups (e.g. domain transfer) using several public and in-house datasets. Our results show that deep features generally outperform other types of features when there is no domain shift and the new task is closely related to the one used to train the model. However, when a large domain or task shift is present, the Fisher-Vector shallow features generalize better and often obtain the best results.

David Novotný, Diane Larlus, Florent Perronnin et al.

Fisher Vectors and related orderless visual statistics have demonstrated excellent performance in object detection, sometimes superior to established approaches such as the Deformable Part Models. However, it remains unclear how these models can capture complex appearance variations using visual codebooks of limited sizes and coarse geometric information. In this work, we propose to interpret Fisher-Vector-based object detectors as part-based models. Through the use of several visualizations and experiments, we show that this is a useful insight to explain the good performance of the model. Furthermore, we reveal for the first time several interesting properties of the FV, including its ability to work well using only a small subset of input patches and visual words. Finally, we discuss the relation of the FV and DPM detectors, pointing out differences and commonalities between them.

Yangmuzi Zhang, Diane Larlus, Florent Perronnin

A natural approach to teaching a visual concept, e.g. a bird species, is to show relevant images. However, not all relevant images represent a concept equally well. In other words, they are not necessarily iconic. This observation raises three questions. Is iconicity a subjective property? If not, can we predict iconicity? And what exactly makes an image iconic? We provide answers to these questions through an extensive experimental study on a challenging fine-grained dataset of birds. We first show that iconicity ratings are consistent across individuals, even when they are not domain experts, thus demonstrating that iconicity is not purely subjective. We then consider an exhaustive list of properties that are intuitively related to iconicity and measure their correlation with these iconicity ratings. We combine them to predict iconicity of new unseen images. We also propose a direct iconicity predictor that is discriminatively trained with iconicity ratings. By combining both systems, we get an iconicity prediction that approaches human performance.

Neda Salamati, Diane Larlus, Gabriela Csurka et al.

Recent progress in computational photography has shown that we can acquire near-infrared (NIR) information in addition to the normal visible (RGB) band, with only slight modifications to standard digital cameras. Due to the proximity of the NIR band to visible radiation, NIR images share many properties with visible images. However, as a result of the material dependent reflection in the NIR part of the spectrum, such images reveal different characteristics of the scene. We investigate how to effectively exploit these differences to improve performance on the semantic image segmentation task. Based on a state-of-the-art segmentation framework and a novel manually segmented image database (both indoor and outdoor scenes) that contain 4-channel images (RGB+NIR), we study how to best incorporate the specific characteristics of the NIR response. We show that adding NIR leads to improved performance for classes that correspond to a specific type of material in both outdoor and indoor scenes. We also discuss the results with respect to the physical properties of the NIR response.