Eduardo Valle

Zahra Mirikharaji, Kumar Abhishek, Alceu Bissoto et al.

Skin cancer is a major public health problem that could benefit from computer-aided diagnosis to reduce the burden of this common disease. Skin lesion segmentation from images is an important step toward achieving this goal. However, the presence of natural and artificial artifacts (e.g., hair and air bubbles), intrinsic factors (e.g., lesion shape and contrast), and variations in image acquisition conditions make skin lesion segmentation a challenging task. Recently, various researchers have explored the applicability of deep learning models to skin lesion segmentation. In this survey, we cross-examine 177 research papers that deal with deep learning-based segmentation of skin lesions. We analyze these works along several dimensions, including input data (datasets, preprocessing, and synthetic data generation), model design (architecture, modules, and losses), and evaluation aspects (data annotation requirements and segmentation performance). We discuss these dimensions both from the viewpoint of select seminal works, and from a systematic viewpoint, examining how those choices have influenced current trends, and how their limitations should be addressed. To facilitate comparisons, we summarize all examined works in a comprehensive table as well as an interactive table available online at https://github.com/sfu-mial/skin-lesion-segmentation-survey.

Alceu Bissoto, Catarina Barata, Eduardo Valle et al.

Deep Learning failure cases are abundant, particularly in the medical area. Recent studies in out-of-distribution generalization have advanced considerably on well-controlled synthetic datasets, but they do not represent medical imaging contexts. We propose a pipeline that relies on artifacts annotation to enable generalization evaluation and debiasing for the challenging skin lesion analysis context. First, we partition the data into levels of increasingly higher biased training and test sets for better generalization assessment. Then, we create environments based on skin lesion artifacts to enable domain generalization methods. Finally, after robust training, we perform a test-time debiasing procedure, reducing spurious features in inference images. Our experiments show our pipeline improves performance metrics in biased cases, and avoids artifacts when using explanation methods. Still, when evaluating such models in out-of-distribution data, they did not prefer clinically-meaningful features. Instead, performance only improved in test sets that present similar artifacts from training, suggesting models learned to ignore the known set of artifacts. Our results raise a concern that debiasing models towards a single aspect may not be enough for fair skin lesion analysis.

Nermin Samet, Cédric Rommel, David Picard et al.

We introduce a novel approach for 3D whole-body pose estimation, addressing the challenge of scale -- and deformability -- variance across body parts brought by the challenge of extending the 17 major joints on the human body to fine-grained keypoints on the face and hands. In addition to addressing the challenge of exploiting motion in unevenly sampled data, we combine stable diffusion to a hierarchical part representation which predicts the relative locations of fine-grained keypoints within each part (e.g., face) with respect to the part's local reference frame. On the H3WB dataset, our method greatly outperforms the current state of the art, which fails to exploit the temporal information. We also show considerable improvements compared to other spatiotemporal 3D human-pose estimation approaches that fail to account for the body part specificities. Code is available at https://github.com/valeoai/PAFUSE.

Levy Chaves, Alceu Bissoto, Eduardo Valle et al.

Transfer learning boosts the performance of medical image analysis by enabling deep learning (DL) on small datasets through the knowledge acquired from large ones. As the number of DL architectures explodes, exhaustively attempting all candidates becomes unfeasible, motivating cheaper alternatives for choosing them. Transferability scoring methods emerge as an enticing solution, allowing to efficiently calculate a score that correlates with the architecture accuracy on any target dataset. However, since transferability scores have not been evaluated on medical datasets, their use in this context remains uncertain, preventing them from benefiting practitioners. We fill that gap in this work, thoroughly evaluating seven transferability scores in three medical applications, including out-of-distribution scenarios. Despite promising results in general-purpose datasets, our results show that no transferability score can reliably and consistently estimate target performance in medical contexts, inviting further work in that direction.

Juliano S. Assine, J. C. S. Santos Filho, Eduardo Valle et al.

The execution of large deep neural networks (DNN) at mobile edge devices requires considerable consumption of critical resources, such as energy, while imposing demands on hardware capabilities. In approaches based on edge computing the execution of the models is offloaded to a compute-capable device positioned at the edge of 5G infrastructures. The main issue of the latter class of approaches is the need to transport information-rich signals over wireless links with limited and time-varying capacity. The recent split computing paradigm attempts to resolve this impasse by distributing the execution of DNN models across the layers of the systems to reduce the amount of data to be transmitted while imposing minimal computing load on mobile devices. In this context, we propose a novel split computing approach based on slimmable ensemble encoders. The key advantage of our design is the ability to adapt computational load and transmitted data size in real-time with minimal overhead and time. This is in contrast with existing approaches, where the same adaptation requires costly context switching and model loading. Moreover, our model outperforms existing solutions in terms of compression efficacy and execution time, especially in the context of weak mobile devices. We present a comprehensive comparison with the most advanced split computing solutions, as well as an experimental evaluation on GPU-less devices.

Tuan-Hung Vu, Eduardo Valle, Andrei Bursuc et al.

We propose the unified BRAVO challenge to benchmark the reliability of semantic segmentation models under realistic perturbations and unknown out-of-distribution (OOD) scenarios. We define two categories of reliability: (1) semantic reliability, which reflects the model's accuracy and calibration when exposed to various perturbations; and (2) OOD reliability, which measures the model's ability to detect object classes that are unknown during training. The challenge attracted nearly 100 submissions from international teams representing notable research institutions. The results reveal interesting insights into the importance of large-scale pre-training and minimal architectural design in developing robust and reliable semantic segmentation models.

Alceu Bissoto, Catarina Barata, Eduardo Valle et al.

Skin lesion analysis models are biased by artifacts placed during image acquisition, which influence model predictions despite carrying no clinical information. Solutions that address this problem by regularizing models to prevent learning those spurious features achieve only partial success, and existing test-time debiasing techniques are inappropriate for skin lesion analysis due to either making unrealistic assumptions on the distribution of test data or requiring laborious annotation from medical practitioners. We propose TTS (Test-Time Selection), a human-in-the-loop method that leverages positive (e.g., lesion area) and negative (e.g., artifacts) keypoints in test samples. TTS effectively steers models away from exploiting spurious artifact-related correlations without retraining, and with less annotation requirements. Our solution is robust to a varying availability of annotations, and different levels of bias. We showcase on the ISIC2019 dataset (for which we release a subset of annotated images) how our model could be deployed in the real-world for mitigating bias.

Victor Besnier, Mickael Chen, David Hurych et al.

Masked Generative Image Transformers (MaskGIT) have emerged as a scalable and efficient image generation framework, able to deliver high-quality visuals with low inference costs. However, MaskGIT's token unmasking scheduler, an essential component of the framework, has not received the attention it deserves. We analyze the sampling objective in MaskGIT, based on the mutual information between tokens, and elucidate its shortcomings. We then propose a new sampling strategy based on our Halton scheduler instead of the original Confidence scheduler. More precisely, our method selects the token's position according to a quasi-random, low-discrepancy Halton sequence. Intuitively, that method spreads the tokens spatially, progressively covering the image uniformly at each step. Our analysis shows that it allows reducing non-recoverable sampling errors, leading to simpler hyper-parameters tuning and better quality images. Our scheduler does not require retraining or noise injection and may serve as a simple drop-in replacement for the original sampling strategy. Evaluation of both class-to-image synthesis on ImageNet and text-to-image generation on the COCO dataset demonstrates that the Halton scheduler outperforms the Confidence scheduler quantitatively by reducing the FID and qualitatively by generating more diverse and more detailed images. Our code is at https://github.com/valeoai/Halton-MaskGIT.

Levy Chaves, Chao Zhou, Rebekka Burkholz et al.

Model merging integrates multiple task-specific models into a single consolidated one. Recent research has made progress in improving merging performance for in-distribution or multi-task scenarios, but domain generalization in model merging remains underexplored. We investigate how merging models fine-tuned on distinct domains affects generalization to unseen domains. Through an analysis of parameter competition in the task matrix using singular value decomposition, we show that merging models trained under different distribution shifts induces stronger conflicts between their subspaces compared to traditional multi-task settings. To mitigate this issue, we propose SCORE (Subspace COnflict-Resolving mErging), a method designed to alleviate such singular subspace conflicts. SCORE finds a shared orthogonal basis by computing the principal components of the concatenated leading singular vectors of all models. It then projects each task matrix into the shared basis, pruning off-diagonal components to remove conflicting singular directions. SCORE consistently outperforms, on average, existing model merging approaches in domain generalization settings across a variety of architectures and model scales, demonstrating its effectiveness and scalability.

Florent Bartoccioni, Elias Ramzi, Victor Besnier et al.

We explore the potential of large-scale generative video models for autonomous driving, introducing an open-source auto-regressive video model (VaViM) and its companion video-action model (VaVAM) to investigate how video pre-training transfers to real-world driving. VaViM is a simple auto-regressive video model that predicts frames using spatio-temporal token sequences. We show that it captures the semantics and dynamics of driving scenes. VaVAM, the video-action model, leverages the learned representations of VaViM to generate driving trajectories through imitation learning. Together, the models form a complete perception-to-action pipeline. We evaluate our models in open- and closed-loop driving scenarios, revealing that video-based pre-training holds promise for autonomous driving. Key insights include the semantic richness of the learned representations, the benefits of scaling for video synthesis, and the complex relationship between model size, data, and safety metrics in closed-loop evaluations. We release code and model weights at https://github.com/valeoai/VideoActionModel

Salma Galaaoui, Eduardo Valle, David Picard et al.

In this paper, we present a comprehensive review of 3D human pose estimation and human mesh recovery from in-the-wild LiDAR point clouds. We compare existing approaches across several key dimensions, and propose a structured taxonomy to classify these methods. Following this taxonomy, we analyze each method's strengths, limitations, and design choices. In addition, (i) we perform a quantitative comparison of the three most widely used datasets, detailing their characteristics; (ii) we compile unified definitions of all evaluation metrics; and (iii) we establish benchmark tables for both tasks on these datasets to enable fair comparisons and promote progress in the field. We also outline open challenges and research directions critical for advancing LiDAR-based 3D human understanding. Moreover, we maintain an accompanying webpage that organizes papers according to our taxonomy and continuously update it with new studies: https://github.com/valeoai/3D-Human-Pose-Shape-Estimation-from-LiDAR

Yihong Xu, Éloi Zablocki, Alexandre Boulch et al.

Motion forecasting is crucial in autonomous driving systems to anticipate the future trajectories of surrounding agents such as pedestrians, vehicles, and traffic signals. In end-to-end forecasting, the model must jointly detect and track from sensor data (cameras or LiDARs) the past trajectories of the different elements of the scene and predict their future locations. We depart from the current trend of tackling this task via end-to-end training from perception to forecasting, and instead use a modular approach. We individually build and train detection, tracking and forecasting modules. We then only use consecutive finetuning steps to integrate the modules better and alleviate compounding errors. We conduct an in-depth study on the finetuning strategies and it reveals that our simple yet effective approach significantly improves performance on the end-to-end forecasting benchmark. Consequently, our solution ranks first in the Argoverse 2 End-to-end Forecasting Challenge, with 63.82 mAPf. We surpass forecasting results by +17.1 points over last year's winner and by +13.3 points over this year's runner-up. This remarkable performance in forecasting can be explained by our modular paradigm, which integrates finetuning strategies and significantly outperforms the end-to-end-trained counterparts. The code, model weights and results are made available https://github.com/valeoai/valeo4cast.

Éloi Zablocki, Valentin Gerard, Amaia Cardiel et al.

Understanding deep models is crucial for deploying them in safety-critical applications. We introduce GIFT, a framework for deriving post-hoc, global, interpretable, and faithful textual explanations for vision classifiers. GIFT starts from local faithful visual counterfactual explanations and employs (vision) language models to translate those into global textual explanations. Crucially, GIFT provides a verification stage measuring the causal effect of the proposed explanations on the classifier decision. Through experiments across diverse datasets, including CLEVR, CelebA, and BDD, we demonstrate that GIFT effectively reveals meaningful insights, uncovering tasks, concepts, and biases used by deep vision classifiers. The framework is released at https://github.com/valeoai/GIFT.

Arthur Douillard, Matthieu Cord, Charles Ollion et al.

Lifelong learning has attracted much attention, but existing works still struggle to fight catastrophic forgetting and accumulate knowledge over long stretches of incremental learning. In this work, we propose PODNet, a model inspired by representation learning. By carefully balancing the compromise between remembering the old classes and learning new ones, PODNet fights catastrophic forgetting, even over very long runs of small incremental tasks --a setting so far unexplored by current works. PODNet innovates on existing art with an efficient spatial-based distillation-loss applied throughout the model and a representation comprising multiple proxy vectors for each class. We validate those innovations thoroughly, comparing PODNet with three state-of-the-art models on three datasets: CIFAR100, ImageNet100, and ImageNet1000. Our results showcase a significant advantage of PODNet over existing art, with accuracy gains of 12.10, 6.51, and 2.85 percentage points, respectively. Code is available at https://github.com/arthurdouillard/incremental_learning.pytorch

Cédric Rommel, Victor Letzelter, Nermin Samet et al.

We propose ManiPose, a manifold-constrained multi-hypothesis model for human-pose 2D-to-3D lifting. We provide theoretical and empirical evidence that, due to the depth ambiguity inherent to monocular 3D human pose estimation, traditional regression models suffer from pose-topology consistency issues, which standard evaluation metrics (MPJPE, P-MPJPE and PCK) fail to assess. ManiPose addresses depth ambiguity by proposing multiple candidate 3D poses for each 2D input, each with its estimated plausibility. Unlike previous multi-hypothesis approaches, ManiPose forgoes generative models, greatly facilitating its training and usage. By constraining the outputs to lie on the human pose manifold, ManiPose guarantees the consistency of all hypothetical poses, in contrast to previous works. We showcase the performance of ManiPose on real-world datasets, where it outperforms state-of-the-art models in pose consistency by a large margin while being very competitive on the MPJPE metric.

Yuanchao Wang, Tian Qin, Eduardo Valle et al.

Out-of-distribution (OOD) detection is critical for deploying image classifiers in safety-sensitive environments, yet existing detectors often struggle when OOD samples are semantically similar to the in-distribution (ID) classes. We present BootOOD, a fully self-supervised OOD detection framework that bootstraps exclusively from ID data and is explicitly designed to handle semantically challenging OOD samples. BootOOD synthesizes pseudo-OOD features through simple transformations of ID representations and leverages Neural Collapse (NC), where ID features cluster tightly around class means with consistent feature norms. Unlike prior approaches that aim to constrain OOD features into subspaces orthogonal to the collapsed ID means, BootOOD introduces a lightweight auxiliary head that performs radius-based classification on feature norms. This design decouples OOD detection from the primary classifier and imposes a relaxed requirement: OOD samples are learned to have smaller feature norms than ID features, which is easier to satisfy when ID and OOD are semantically close. Experiments on CIFAR-10, CIFAR-100, and ImageNet-200 show that BootOOD outperforms prior post-hoc methods, surpasses training-based methods without outlier exposure, and is competitive with state-of-the-art outlier-exposure approaches while maintaining or improving ID accuracy.

Levy Chaves, Eduardo Valle, Sandra Avila

Model merging provides a cost-effective and data-efficient combination of specialized deep neural networks through parameter integration. This technique leverages expert models across downstream tasks without requiring retraining. Most model merging approaches critically depend on scaling hyper-parameters $λ$, which weight each model's contribution globally or individually. Principled approaches for setting scaling factors without accessing any data (data-free) are scarce, often leading researchers to tune $λ$ using privileged data from the evaluation set, which is obviously unfeasible in practice. To address this limitation, we introduce Weight Weaving, a plug-and-play technique that pools model weights across $λ$ values search space using user-defined pooling functions, such as averaging, random selection, or even existing model merging methods. Our method demonstrates high modularity, imposing minimal constraints on the search space. It operates orthogonally to existing model merging methods and eliminates evaluation data requirements. We validate Weight Weaving across three ViT variants in three experimental setups: vision multi-task learning, vision continual learning, and domain generalization. Our method consistently improves the performance of several model merging methods, achieving average accuracy gains of up to 15.9 percentage points in a data-free setting.

Victor Besnier, David Hurych, Andrei Bursuc et al.

Recent advances in image and video generation have raised significant interest from both academia and industry. A key challenge in this field is improving inference efficiency, as model size and the number of inference steps directly impact the commercial viability of generative models while also posing fundamental scientific challenges. A promising direction involves combining auto-regressive sequential token modeling with multi-token prediction per step, reducing inference time by up to an order of magnitude. However, predicting multiple tokens in parallel can introduce structural inconsistencies due to token incompatibilities, as capturing complex joint dependencies during training remains challenging. Traditionally, once tokens are sampled, there is no mechanism to backtrack and refine erroneous predictions. We propose a method for self-correcting image generation by iteratively refining sampled tokens. We achieve this with a novel training scheme that injects random tokens in the context, improving robustness and enabling token fixing during sampling. Our method preserves the efficiency benefits of parallel token prediction while significantly enhancing generation quality. We evaluate our approach on image generation using the ImageNet-256 and CIFAR-10 datasets, as well as on video generation with UCF-101 and NuScenes, demonstrating substantial improvements across both modalities.

Cédric Rommel, Eduardo Valle, Mickaël Chen et al.

We present an innovative approach to 3D Human Pose Estimation (3D-HPE) by integrating cutting-edge diffusion models, which have revolutionized diverse fields, but are relatively unexplored in 3D-HPE. We show that diffusion models enhance the accuracy, robustness, and coherence of human pose estimations. We introduce DiffHPE, a novel strategy for harnessing diffusion models in 3D-HPE, and demonstrate its ability to refine standard supervised 3D-HPE. We also show how diffusion models lead to more robust estimations in the face of occlusions, and improve the time-coherence and the sagittal symmetry of predictions. Using the Human\,3.6M dataset, we illustrate the effectiveness of our approach and its superiority over existing models, even under adverse situations where the occlusion patterns in training do not match those in inference. Our findings indicate that while standalone diffusion models provide commendable performance, their accuracy is even better in combination with supervised models, opening exciting new avenues for 3D-HPE research.

Alceu Bissoto, Catarina Barata, Eduardo Valle et al.

Distribution shifts are common in real-world datasets and can affect the performance and reliability of deep learning models. In this paper, we study two types of distribution shifts: diversity shifts, which occur when test samples exhibit patterns unseen during training, and correlation shifts, which occur when test data present a different correlation between seen invariant and spurious features. We propose an integrated protocol to analyze both types of shifts using datasets where they co-exist in a controllable manner. Finally, we apply our approach to a real-world classification problem of skin cancer analysis, using out-of-distribution datasets and specialized bias annotations. Our protocol reveals three findings: 1) Models learn and propagate correlation shifts even with low-bias training; this poses a risk of accumulating and combining unaccountable weak biases; 2) Models learn robust features in high- and low-bias scenarios but use spurious ones if test samples have them; this suggests that spurious correlations do not impair the learning of robust features; 3) Diversity shift can reduce the reliance on spurious correlations; this is counter intuitive since we expect biased models to depend more on biases when invariant features are missing. Our work has implications for distribution shift research and practice, providing new insights into how models learn and rely on spurious correlations under different types of shifts.

Levy Chaves, Alceu Bissoto, Eduardo Valle et al.

Self-supervised pre-training appears as an advantageous alternative to supervised pre-trained for transfer learning. By synthesizing annotations on pretext tasks, self-supervision allows to pre-train models on large amounts of pseudo-labels before fine-tuning them on the target task. In this work, we assess self-supervision for the diagnosis of skin lesions, comparing three self-supervised pipelines to a challenging supervised baseline, on five test datasets comprising in- and out-of-distribution samples. Our results show that self-supervision is competitive both in improving accuracies and in reducing the variability of outcomes. Self-supervision proves particularly useful for low training data scenarios ($<1\,500$ and $<150$ samples), where its ability to stabilize the outcomes is essential to provide sound results.

Juliano S. Assine, J. C. S. Santos Filho, Eduardo Valle

In the past few years, mobile deep-learning deployment progressed by leaps and bounds, but solutions still struggle to accommodate its severe and fluctuating operational restrictions, which include bandwidth, latency, computation, and energy. In this work, we help to bridge that gap, introducing the first configurable solution for object detection that manages the triple communication-computation-accuracy trade-off with a single set of weights. Our solution shows state-of-the-art results on COCO-2017, adding only a minor penalty on the base EfficientDet-D2 architecture. Our design is robust to the choice of base architecture and compressor and should adapt well for future architectures.

Alceu Bissoto, Eduardo Valle, Sandra Avila

Despite the growing availability of high-quality public datasets, the lack of training samples is still one of the main challenges of deep-learning for skin lesion analysis. Generative Adversarial Networks (GANs) appear as an enticing alternative to alleviate the issue, by synthesizing samples indistinguishable from real images, with a plethora of works employing them for medical applications. Nevertheless, carefully designed experiments for skin-lesion diagnosis with GAN-based data augmentation show favorable results only on out-of-distribution test sets. For GAN-based data anonymization $-$ where the synthetic images replace the real ones $-$ favorable results also only appear for out-of-distribution test sets. Because of the costs and risks associated with GAN usage, those results suggest caution in their adoption for medical applications.

Arthur Douillard, Eduardo Valle, Charles Ollion et al.

Continual learning aims to learn tasks sequentially, with (often severe) constraints on the storage of old learning samples, without suffering from catastrophic forgetting. In this work, we propose prescient continual learning, a novel experimental setting, to incorporate existing information about the classes, prior to any training data. Usually, each task in a traditional continual learning setting evaluates the model on present and past classes, the latter with a limited number of training samples. Our setting adds future classes, with no training samples at all. We introduce Ghost Model, a representation-learning-based model for continual learning using ideas from zero-shot learning. A generative model of the representation space in concert with a careful adjustment of the losses allows us to exploit insights from future classes to constraint the spatial arrangement of the past and current classes. Quantitative results on the AwA2 and aP\&Y datasets and detailed visualizations showcase the interest of this new setting and the method we propose to address it.

Vinicius Ribeiro, Sandra Avila, Eduardo Valle

Segmenting skin lesions images is relevant both for itself and for assisting in lesion classification, but suffers from the challenge in obtaining annotated data. In this work, we show that segmentation may improve with less data, by selecting the training samples with best inter-annotator agreement, and conditioning the ground-truth masks to remove excessive detail. We perform an exhaustive experimental design considering several sources of variation, including three different test sets, two different deep-learning architectures, and several replications, for a total of 540 experimental runs. We found that sample selection and detail removal may have impacts corresponding, respectively, to 12% and 16% of the one obtained by picking a better deep-learning model.

Alceu Bissoto, Eduardo Valle, Sandra Avila

Data-driven models are now deployed in a plethora of real-world applications - including automated diagnosis - but models learned from data risk learning biases from that same data. When models learn spurious correlations not found in real-world situations, their deployment for critical tasks, such as medical decisions, can be catastrophic. In this work we address this issue for skin-lesion classification models, with two objectives: finding out what are the spurious correlations exploited by biased networks, and debiasing the models by removing such spurious correlations from them. We perform a systematic integrated analysis of 7 visual artifacts (which are possible sources of biases exploitable by networks), employ a state-of-the-art technique to prevent the models from learning spurious correlations, and propose datasets to test models for the presence of bias. We find out that, despite interesting results that point to promising future research, current debiasing methods are not ready to solve the bias issue for skin-lesion models.

Juliano S. Assine, Alan Godoy, Eduardo Valle

Despite recent advances in architectures for mobile devices, deep learning computational requirements remains prohibitive for most embedded devices. To address that issue, we envision sharing the computational costs of inference between local devices and the cloud, taking advantage of the compression performed by the first layers of the networks to reduce communication costs. Inference in such distributed setting would allow new applications, but requires balancing a triple trade-off between computation cost, communication bandwidth, and model accuracy. We explore that trade-off by studying the compressibility of representations at different stages of MobileNetV2, showing those results agree with theoretical intuitions about deep learning, and that an optimal splitting layer for network can be found with a simple PCA-based compression scheme.

Alceu Bissoto, Eduardo Valle, Sandra Avila

Generative Adversarial Networks fostered a newfound interest in generative models, resulting in a swelling wave of new works that new-coming researchers may find formidable to surf. In this paper, we intend to help those researchers, by splitting that incoming wave into six "fronts": Architectural Contributions, Conditional Techniques, Normalization and Constraint Contributions, Loss Functions, Image-to-image Translations, and Validation Metrics. The division in fronts organizes literature into approachable blocks, ultimately communicating to the reader how the area is evolving. Previous surveys in the area, which this works also tabulates, focus on a few of those fronts, leaving a gap that we propose to fill with a more integrated, comprehensive overview. Here, instead of an exhaustive survey, we opt for a straightforward review: our target is to be an entry point to this vast literature, and also to be able to update experienced researchers to the newest techniques.

Vinicius Ribeiro, Sandra Avila, Eduardo Valle

In this work, we explore the issue of the inter-annotator agreement for training and evaluating automated segmentation of skin lesions. We explore what different degrees of agreement represent, and how they affect different use cases for segmentation. We also evaluate how conditioning the ground truths using different (but very simple) algorithms may help to enhance agreement and may be appropriate for some use cases. The segmentation of skin lesions is a cornerstone task for automated skin lesion analysis, useful both as an end-result to locate/detect the lesions and as an ancillary task for lesion classification. Lesion segmentation, however, is a very challenging task, due not only to the challenge of image segmentation itself but also to the difficulty in obtaining properly annotated data. Detecting accurately the borders of lesions is challenging even for trained humans, since, for many lesions, those borders are fuzzy and ill-defined. Using lesions and annotations from the ISIC Archive, we estimate inter-annotator agreement for skin-lesion segmentation and propose several simple procedures that may help to improve inter-annotator agreement if used to condition the ground truths.

Fábio Perez, Sandra Avila, Eduardo Valle

Convolutional neural networks (CNNs) deliver exceptional results for computer vision, including medical image analysis. With the growing number of available architectures, picking one over another is far from obvious. Existing art suggests that, when performing transfer learning, the performance of CNN architectures on ImageNet correlates strongly with their performance on target tasks. We evaluate that claim for melanoma classification, over 9 CNNs architectures, in 5 sets of splits created on the ISIC Challenge 2017 dataset, and 3 repeated measures, resulting in 135 models. The correlations we found were, to begin with, much smaller than those reported by existing art, and disappeared altogether when we considered only the top-performing networks: uncontrolled nuisances (i.e., splits and randomness) overcome any of the analyzed factors. Whenever possible, the best approach for melanoma classification is still to create ensembles of multiple models. We compared two choices for selecting which models to ensemble: picking them at random (among a pool of high-quality ones) vs. using the validation set to determine which ones to pick first. For small ensembles, we found a slight advantage on the second approach but found that random choice was also competitive. Although our aim in this paper was not to maximize performance, we easily reached AUCs comparable to the first place on the ISIC Challenge 2017.

Alceu Bissoto, Michel Fornaciali, Eduardo Valle et al.

Melanoma is the deadliest form of skin cancer. Automated skin lesion analysis plays an important role for early detection. Nowadays, the ISIC Archive and the Atlas of Dermoscopy dataset are the most employed skin lesion sources to benchmark deep-learning based tools. However, all datasets contain biases, often unintentional, due to how they were acquired and annotated. Those biases distort the performance of machine-learning models, creating spurious correlations that the models can unfairly exploit, or, contrarily destroying cogent correlations that the models could learn. In this paper, we propose a set of experiments that reveal both types of biases, positive and negative, in existing skin lesion datasets. Our results show that models can correctly classify skin lesion images without clinically-meaningful information: disturbingly, the machine-learning model learned over images where no information about the lesion remains, presents an accuracy above the AI benchmark curated with dermatologists' performances. That strongly suggests spurious correlations guiding the models. We fed models with additional clinically meaningful information, which failed to improve the results even slightly, suggesting the destruction of cogent correlations. Our main findings raise awareness of the limitations of models trained and evaluated in small datasets such as the ones we evaluated, and may suggest future guidelines for models intended for real-world deployment.

Alceu Bissoto, Fábio Perez, Eduardo Valle et al.

Skin cancer is by far the most common type of cancer. Early detection is the key to increase the chances for successful treatment significantly. Currently, Deep Neural Networks are the state-of-the-art results on automated skin cancer classification. To push the results further, we need to address the lack of annotated data, which is expensive and require much effort from specialists. To bypass this problem, we propose using Generative Adversarial Networks for generating realistic synthetic skin lesion images. To the best of our knowledge, our results are the first to show visually-appealing synthetic images that comprise clinically-meaningful information.

Fábio Perez, Cristina Vasconcelos, Sandra Avila et al.

Deep learning models show remarkable results in automated skin lesion analysis. However, these models demand considerable amounts of data, while the availability of annotated skin lesion images is often limited. Data augmentation can expand the training dataset by transforming input images. In this work, we investigate the impact of 13 data augmentation scenarios for melanoma classification trained on three CNNs (Inception-v4, ResNet, and DenseNet). Scenarios include traditional color and geometric transforms, and more unusual augmentations such as elastic transforms, random erasing and a novel augmentation that mixes different lesions. We also explore the use of data augmentation at test-time and the impact of data augmentation on various dataset sizes. Our results confirm the importance of data augmentation in both training and testing and show that it can lead to more performance gains than obtaining new images. The best scenario results in an AUC of 0.882 for melanoma classification without using external data, outperforming the top-ranked submission (0.874) for the ISIC Challenge 2017, which was trained with additional data.

Alceu Bissoto, Fábio Perez, Vinícius Ribeiro et al.

This extended abstract describes the participation of RECOD Titans in parts 1 to 3 of the ISIC Challenge 2018 "Skin Lesion Analysis Towards Melanoma Detection" (MICCAI 2018). Although our team has a long experience with melanoma classification and moderate experience with lesion segmentation, the ISIC Challenge 2018 was the very first time we worked on lesion attribute detection. For each task we submitted 3 different ensemble approaches, varying combinations of models and datasets. Our best results on the official testing set, regarding the official metric of each task, were: 0.728 (segmentation), 0.344 (attribute detection) and 0.803 (classification). Those submissions reached, respectively, the 56th, 14th and 9th places.

George Gondim-Ribeiro, Pedro Tabacof, Eduardo Valle

Adversarial attacks are malicious inputs that derail machine-learning models. We propose a scheme to attack autoencoders, as well as a quantitative evaluation framework that correlates well with the qualitative assessment of the attacks. We assess --- with statistically validated experiments --- the resistance to attacks of three variational autoencoders (simple, convolutional, and DRAW) in three datasets (MNIST, SVHN, CelebA), showing that both DRAW's recurrence and attention mechanism lead to better resistance. As autoencoders are proposed for compressing data --- a scenario in which their safety is paramount --- we expect more attention will be given to adversarial attacks on them.

Eduardo Valle, Michel Fornaciali, Afonso Menegola et al.

Deep learning fostered a leap ahead in automated skin lesion analysis in the last two years. Those models are expensive to train and difficult to parameterize. Objective: We investigate methodological issues for designing and evaluating deep learning models for skin lesion analysis. We explore 10 choices faced by researchers: use of transfer learning, model architecture, train dataset, image resolution, type of data augmentation, input normalization, use of segmentation, duration of training, additional use of SVMs, and test data augmentation. Methods: We perform two full factorial experiments, for five different test datasets, resulting in 2560 exhaustive trials in our main experiment, and 1280 trials in our assessment of transfer learning. We analyze both with multi-way ANOVA. We use the exhaustive trials to simulate sequential decisions and ensembles, with and without the use of privileged information from the test set. Results -- main experiment: Amount of train data has disproportionate influence, explaining almost half the variation in performance. Of the other factors, test data augmentation and input resolution are the most influential. Deeper models, when combined, with extra data, also help. -- transfer experiment: Transfer learning is critical, its absence brings huge performance penalties. -- simulations: Ensembles of models are the best option to provide reliable results with limited resources, without using privileged information and sacrificing methodological rigor. Conclusions and Significance: Advancing research on automated skin lesion analysis requires curating larger public datasets. Indirect use of privileged information from the test set to design the models is a subtle, but frequent methodological mistake that leads to overoptimistic results. Ensembles of models are a cost-effective alternative to the expensive full-factorial and to the unstable sequential designs.

Afonso Menegola, Michel Fornaciali, Ramon Pires et al.

Knowledge transfer impacts the performance of deep learning -- the state of the art for image classification tasks, including automated melanoma screening. Deep learning's greed for large amounts of training data poses a challenge for medical tasks, which we can alleviate by recycling knowledge from models trained on different tasks, in a scheme called transfer learning. Although much of the best art on automated melanoma screening employs some form of transfer learning, a systematic evaluation was missing. Here we investigate the presence of transfer, from which task the transfer is sourced, and the application of fine tuning (i.e., retraining of the deep learning model after transfer). We also test the impact of picking deeper (and more expensive) models. Our results favor deeper models, pre-trained over ImageNet, with fine-tuning, reaching an AUC of 80.7% and 84.5% for the two skin-lesion datasets evaluated.

Afonso Menegola, Julia Tavares, Michel Fornaciali et al.

This extended abstract describes the participation of RECOD Titans in parts 1 and 3 of the ISIC Challenge 2017 "Skin Lesion Analysis Towards Melanoma Detection" (ISBI 2017). Although our team has a long experience with melanoma classification, the ISIC Challenge 2017 was the very first time we worked on skin-lesion segmentation. For part 1 (segmentation), our final submission used four of our models: two trained with all 2000 samples, without a validation split, for 250 and for 500 epochs respectively; and other two trained and validated with two different 1600/400 splits, for 220 epochs. Those four models, individually, achieved between 0.780 and 0.783 official validation scores. Our final submission averaged the output of those four models achieved a score of 0.793. For part 3 (classification), the submitted test run as well as our last official validation run were the result from a meta-model that assembled seven base deep-learning models: three based on Inception-V4 trained on our largest dataset; three based on Inception trained on our smallest dataset; and one based on ResNet-101 trained on our smaller dataset. The results of those component models were stacked in a meta-learning layer based on an SVM trained on the validation set of our largest dataset.

Ramon Oliveira, Pedro Tabacof, Eduardo Valle

We evaluate the uncertainty quality in neural networks using anomaly detection. We extract uncertainty measures (e.g. entropy) from the predictions of candidate models, use those measures as features for an anomaly detector, and gauge how well the detector differentiates known from unknown classes. We assign higher uncertainty quality to candidate models that lead to better detectors. We also propose a novel method for sampling a variational approximation of a Bayesian neural network, called One-Sample Bayesian Approximation (OSBA). We experiment on two datasets, MNIST and CIFAR10. We compare the following candidate neural network models: Maximum Likelihood, Bayesian Dropout, OSBA, and --- for MNIST --- the standard variational approximation. We show that Bayesian Dropout and OSBA provide better uncertainty information than Maximum Likelihood, and are essentially equivalent to the standard variational approximation, but much faster.

Pedro Tabacof, Julia Tavares, Eduardo Valle

We investigate adversarial attacks for autoencoders. We propose a procedure that distorts the input image to mislead the autoencoder in reconstructing a completely different target image. We attack the internal latent representations, attempting to make the adversarial input produce an internal representation as similar as possible as the target's. We find that autoencoders are much more robust to the attack than classifiers: while some examples have tolerably small input distortion, and reasonable similarity to the target image, there is a quasi-linear trade-off between those aims. We report results on MNIST and SVHN datasets, and also test regular deterministic autoencoders, reaching similar conclusions in all cases. Finally, we show that the usual adversarial attack for classifiers, while being much easier, also presents a direct proportion between distortion on the input, and misdirection on the output. That proportionality however is hidden by the normalization of the output, which maps a linear layer into non-linear probabilities.

Afonso Menegola, Michel Fornaciali, Ramon Pires et al.

Deep learning is the current bet for image classification. Its greed for huge amounts of annotated data limits its usage in medical imaging context. In this scenario transfer learning appears as a prominent solution. In this report we aim to clarify how transfer learning schemes may influence classification results. We are particularly focused in the automated melanoma screening problem, a case of medical imaging in which transfer learning is still not widely used. We explored transfer with and without fine-tuning, sequential transfers and usage of pre-trained models in general and specific datasets. Although some issues remain open, our findings may drive future researches.

Micael Carvalho, Matthieu Cord, Sandra Avila et al.

In recent years, deep architectures have been used for transfer learning with state-of-the-art performance in many datasets. The properties of their features remain, however, largely unstudied under the transfer perspective. In this work, we present an extensive analysis of the resiliency of feature vectors extracted from deep models, with special focus on the trade-off between performance and compression rate. By introducing perturbations to image descriptions extracted from a deep convolutional neural network, we change their precision and number of dimensions, measuring how it affects the final score. We show that deep features are more robust to these disturbances when compared to classical approaches, achieving a compression rate of 98.4%, while losing only 0.88% of their original score for Pascal VOC 2007.

Michel Fornaciali, Micael Carvalho, Flávia Vasques Bittencourt et al.

In this paper we survey, analyze and criticize current art on automated melanoma screening, reimplementing a baseline technique, and proposing two novel ones. Melanoma, although highly curable when detected early, ends as one of the most dangerous types of cancer, due to delayed diagnosis and treatment. Its incidence is soaring, much faster than the number of trained professionals able to diagnose it. Automated screening appears as an alternative to make the most of those professionals, focusing their time on the patients at risk while safely discharging the other patients. However, the potential of automated melanoma diagnosis is currently unfulfilled, due to the emphasis of current literature on outdated computer vision models. Even more problematic is the irreproducibility of current art. We show how streamlined pipelines based upon current Computer Vision outperform conventional models - a model based on an advanced bags of words reaches an AUC of 84.6%, and a model based on deep neural networks reaches 89.3%, while the baseline (a classical bag of words) stays at 81.2%. We also initiate a dialog to improve reproducibility in our community

Otávio A. B. Penatti, Sandra Avila, Eduardo Valle et al.

Visual dictionaries are a critical component for image classification/retrieval systems based on the bag-of-visual-words (BoVW) model. Dictionaries are usually learned without supervision from a training set of images sampled from the collection of interest. However, for large, general-purpose, dynamic image collections (e.g., the Web), obtaining a representative sample in terms of semantic concepts is not straightforward. In this paper, we evaluate the impact of semantics in the dictionary quality, aiming at verifying the importance of semantic diversity in relation visual diversity for visual dictionaries. In the experiments, we vary the amount of classes used for creating the dictionary and then compute different BoVW descriptors, using multiple codebook sizes and different coding and pooling methods (standard BoVW and Fisher Vectors). Results for image classification show that as visual dictionaries are based on low-level visual appearances, visual diversity is more important than semantic diversity. Our conclusions open the opportunity to alleviate the burden in generating visual dictionaries as we need only a visually diverse set of images instead of the whole collection to create a good dictionary.

Pedro Tabacof, Eduardo Valle

Adversarial examples have raised questions regarding the robustness and security of deep neural networks. In this work we formalize the problem of adversarial images given a pretrained classifier, showing that even in the linear case the resulting optimization problem is nonconvex. We generate adversarial images using shallow and deep classifiers on the MNIST and ImageNet datasets. We probe the pixel space of adversarial images using noise of varying intensity and distribution. We bring novel visualizations that showcase the phenomenon and its high variability. We show that adversarial images appear in large regions in the pixel space, but that, for the same task, a shallow classifier seems more robust to adversarial images than a deep convolutional network.

Thierry Pinheiro Moreira, Mauricio Lisboa Perez, Rafael de Oliveira Werneck et al.

A pet that goes missing is among many people's worst fears: a moment of distraction is enough for a dog or a cat wandering off from home. Some measures help matching lost animals to their owners; but automated visual recognition is one that - although convenient, highly available, and low-cost - is surprisingly overlooked. In this paper, we inaugurate that promising avenue by pursuing face recognition for dogs. We contrast four ready-to-use human facial recognizers (EigenFaces, FisherFaces, LBPH, and a Sparse method) to two original solutions based upon convolutional neural networks: BARK (inspired in architecture-optimized networks employed for human facial recognition) and WOOF (based upon off-the-shelf OverFeat features). Human facial recognizers perform poorly for dogs (up to 60.5% accuracy), showing that dog facial recognition is not a trivial extension of human facial recognition. The convolutional network solutions work much better, with BARK attaining up to 81.1% accuracy, and WOOF, 89.4%. The tests were conducted in two datasets: Flickr-dog, with 42 dogs of two breeds (pugs and huskies); and Snoopybook, with 18 mongrel dogs.

Thiago S. F. X. Teixeira, George Teodoro, Eduardo Valle et al.

Similarity search is critical for many database applications, including the increasingly popular online services for Content-Based Multimedia Retrieval (CBMR). These services, which include image search engines, must handle an overwhelming volume of data, while keeping low response times. Thus, scalability is imperative for similarity search in Web-scale applications, but most existing methods are sequential and target shared-memory machines. Here we address these issues with a distributed, efficient, and scalable index based on Locality-Sensitive Hashing (LSH). LSH is one of the most efficient and popular techniques for similarity search, but its poor referential locality properties has made its implementation a challenging problem. Our solution is based on a widely asynchronous dataflow parallelization with a number of optimizations that include a hierarchical parallelization to decouple indexing and data storage, locality-aware data partition strategies to reduce message passing, and multi-probing to limit memory usage. The proposed parallelization attained an efficiency of 90% in a distributed system with about 800 CPU cores. In particular, the original locality-aware data partition reduced the number of messages exchanged in 30%. Our parallel LSH was evaluated using the largest public dataset for similarity search (to the best of our knowledge) with $10^9$ 128-d SIFT descriptors extracted from Web images. This is two orders of magnitude larger than datasets that previous LSH parallelizations could handle.

George Teodoro, Eduardo Valle, Nathan Mariano et al.

Similarity search in high-dimentional spaces is a pivotal operation found a variety of database applications. Recently, there has been an increase interest in similarity search for online content-based multimedia services. Those services, however, introduce new challenges with respect to the very large volumes of data that have to be indexed/searched, and the need to minimize response times observed by the end-users. Additionally, those users dynamically interact with the systems creating fluctuating query request rates, requiring the search algorithm to adapt in order to better utilize the underline hardware to reduce response times. In order to address these challenges, we introduce hypercurves, a flexible framework for answering approximate k-nearest neighbor (kNN) queries for very large multimedia databases, aiming at online content-based multimedia services. Hypercurves executes on hybrid CPU--GPU environments, and is able to employ those devices cooperatively to support massive query request rates. In order to keep the response times optimal as the request rates vary, it employs a novel dynamic scheduler to partition the work between CPU and GPU. Hypercurves was throughly evaluated using a large database of multimedia descriptors. Its cooperative CPU--GPU execution achieved performance improvements of up to 30x when compared to the single CPU-core version. The dynamic work partition mechanism reduces the observed query response times in about 50% when compared to the best static CPU--GPU task partition configuration. In addition, Hypercurves achieves superlinear scalability in distributed (multi-node) executions, while keeping a high guarantee of equivalence with its sequential version --- thanks to the proof of probabilistic equivalence, which supported its aggressive parallelization design.

Otavio A. B. Penatti, Eduardo Valle, Ricardo da S. Torres

Mid-level features based on visual dictionaries are today a cornerstone of systems for classification and retrieval of images. Those state-of-the-art representations depend crucially on the choice of a codebook (visual dictionary), which is usually derived from the dataset. In general-purpose, dynamic image collections (e.g., the Web), one cannot have the entire collection in order to extract a representative dictionary. However, based on the hypothesis that the dictionary reflects only the diversity of low-level appearances and does not capture semantics, we argue that a dictionary based on a small subset of the data, or even on an entirely different dataset, is able to produce a good representation, provided that the chosen images span a diverse enough portion of the low-level feature space. Our experiments confirm that hypothesis, opening the opportunity to greatly alleviate the burden in generating the codebook, and confirming the feasibility of employing visual dictionaries in large-scale dynamic environments.