Marcos Escudero-Viñolo

Xuanlong Yu, Yi Zuo, Zitao Wang et al. · cmu

This paper outlines the winning solutions employed in addressing the MUAD uncertainty quantification challenge held at ICCV 2023. The challenge was centered around semantic segmentation in urban environments, with a particular focus on natural adversarial scenarios. The report presents the results of 19 submitted entries, with numerous techniques drawing inspiration from cutting-edge uncertainty quantification methodologies presented at prominent conferences in the fields of computer vision and machine learning and journals over the past few years. Within this document, the challenge is introduced, shedding light on its purpose and objectives, which primarily revolved around enhancing the robustness of semantic segmentation in urban scenes under varying natural adversarial conditions. The report then delves into the top-performing solutions. Moreover, the document aims to provide a comprehensive overview of the diverse solutions deployed by all participants. By doing so, it seeks to offer readers a deeper insight into the array of strategies that can be leveraged to effectively handle the inherent uncertainties associated with autonomous driving and semantic segmentation, especially within urban environments.

Elena Luna, Juan Carlos San Miguel, José María Martínez et al.

This letter focuses on the task of Multi-Target Multi-Camera vehicle tracking. We propose to associate single-camera trajectories into multi-camera global trajectories by training a Graph Convolutional Network. Our approach simultaneously processes all cameras providing a global solution, and it is also robust to large cameras unsynchronizations. Furthermore, we design a new loss function to deal with class imbalance. Our proposal outperforms the related work showing better generalization and without requiring ad-hoc manual annotations or thresholds, unlike compared approaches.

Kirill Sirotkin, Pablo Carballeira, Marcos Escudero-Viñolo

Deep neural networks are efficient at learning the data distribution if it is sufficiently sampled. However, they can be strongly biased by non-relevant factors implicitly incorporated in the training data. These include operational biases, such as ineffective or uneven data sampling, but also ethical concerns, as the social biases are implicitly present\textemdash even inadvertently, in the training data or explicitly defined in unfair training schedules. In tasks having impact on human processes, the learning of social biases may produce discriminatory, unethical and untrustworthy consequences. It is often assumed that social biases stem from supervised learning on labelled data, and thus, Self-Supervised Learning (SSL) wrongly appears as an efficient and bias-free solution, as it does not require labelled data. However, it was recently proven that a popular SSL method also incorporates biases. In this paper, we study the biases of a varied set of SSL visual models, trained using ImageNet data, using a method and dataset designed by psychological experts to measure social biases. We show that there is a correlation between the type of the SSL model and the number of biases that it incorporates. Furthermore, the results also suggest that this number does not strictly depend on the model's accuracy and changes throughout the network. Finally, we conclude that a careful SSL model selection process can reduce the number of social biases in the deployed model, whilst keeping high performance.

Roberto Alcover-Couso, Marcos Escudero-Viñolo, Juan C. SanMiguel et al.

In unsupervised domain adaptation (UDA), where models are trained on source data (e.g., synthetic) and adapted to target data (e.g., real-world) without target annotations, addressing the challenge of significant class imbalance remains an open issue. Despite considerable progress in bridging the domain gap, existing methods often experience performance degradation when confronted with highly imbalanced dense prediction visual tasks like semantic and panoptic segmentation. This discrepancy becomes especially pronounced due to the lack of equivalent priors between the source and target domains, turning class imbalanced techniques used for other areas (e.g., image classification) ineffective in UDA scenarios. This paper proposes a class-imbalance mitigation strategy that incorporates class-weights into the UDA learning losses, but with the novelty of estimating these weights dynamically through the loss gradient, defining a Gradient-based class weighting (GBW) learning. GBW naturally increases the contribution of classes whose learning is hindered by large-represented classes, and has the advantage of being able to automatically and quickly adapt to the iteration training outcomes, avoiding explicitly curricular learning patterns common in loss-weighing strategies. Extensive experimentation validates the effectiveness of GBW across architectures (convolutional and transformer), UDA strategies (adversarial, self-training and entropy minimization), tasks (semantic and panoptic segmentation), and datasets (GTA and Synthia). Analysing the source of advantage, GBW consistently increases the recall of low represented classes.

Roberto Alcover-Couso, Juan C. SanMiguel, Marcos Escudero-Viñolo et al.

Merging parameters of multiple models has resurfaced as an effective strategy to enhance task performance and robustness, but prior work is limited by the high costs of ensemble creation and inference. In this paper, we leverage the abundance of freely accessible trained models to introduce a cost-free approach to model merging. It focuses on a layer-wise integration of merged models, aiming to maintain the distinctiveness of the task-specific final layers while unifying the initial layers, which are primarily associated with feature extraction. This approach ensures parameter consistency across all layers, essential for boosting performance. Moreover, it facilitates seamless integration of knowledge, enabling effective merging of models from different datasets and tasks. Specifically, we investigate its applicability in Unsupervised Domain Adaptation (UDA), an unexplored area for model merging, for Semantic and Panoptic Segmentation. Experimental results demonstrate substantial UDA improvements without additional costs for merging same-architecture models from distinct datasets ($\uparrow 2.6\%$ mIoU) and different-architecture models with a shared backbone ($\uparrow 6.8\%$ mIoU). Furthermore, merging Semantic and Panoptic Segmentation models increases mPQ by $\uparrow 7\%$. These findings are validated across a wide variety of UDA strategies, architectures, and datasets.

Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós et al.

Knowledge Distillation (KD) is a strategy for the definition of a set of transferability gangways to improve the efficiency of Convolutional Neural Networks. Feature-based Knowledge Distillation is a subfield of KD that relies on intermediate network representations, either unaltered or depth-reduced via maximum activation maps, as the source knowledge. In this paper, we propose and analyse the use of a 2D frequency transform of the activation maps before transferring them. We pose that\textemdash by using global image cues rather than pixel estimates, this strategy enhances knowledge transferability in tasks such as scene recognition, defined by strong spatial and contextual relationships between multiple and varied concepts. To validate the proposed method, an extensive evaluation of the state-of-the-art in scene recognition is presented. Experimental results provide strong evidences that the proposed strategy enables the student network to better focus on the relevant image areas learnt by the teacher network, hence leading to better descriptive features and higher transferred performance than every other state-of-the-art alternative. We publicly release the training and evaluation framework used along this paper at http://www-vpu.eps.uam.es/publications/DCTBasedKDForSceneRecognition.

Iván de Andrés Tamé, Kirill Sirotkin, Pablo Carballeira et al.

Deep learning technologies have already demonstrated a high potential to build diagnosis support systems from medical imaging data, such as Chest X-Ray images. However, the shortage of labeled data in the medical field represents one key obstacle to narrow down the performance gap with respect to applications in other image domains. In this work, we investigate the benefits of a curricular Self-Supervised Learning (SSL) pretraining scheme with respect to fully-supervised training regimes for pneumonia recognition on Chest X-Ray images of Covid-19 patients. We show that curricular SSL pretraining, which leverages unlabeled data, outperforms models trained from scratch, or pretrained on ImageNet, indicating the potential of performance gains by SSL pretraining on massive unlabeled datasets. Finally, we demonstrate that top-performing SSLpretrained models show a higher degree of attention in the lung regions, embodying models that may be more robust to possible external confounding factors in the training datasets, identified by previous works.

Cecilia Diana-Albelda, Roberto Alcover-Couso, Álvaro García-Martín et al.

Gliomas are aggressive brain tumors that require accurate imaging-based diagnosis, with segmentation playing a critical role in evaluating morphology and treatment decisions. Manual delineation of gliomas is time-consuming and prone to variability, motivating the use of deep learning to improve consistency and alleviate clinical workload. However, existing methods often fail to fully exploit the information available in multi-parametric MRI (mp-MRI), particularly inter-slice contextual features, and typically require considerable computational resources while lacking robustness across tumor type variations. We present GBT-SAM, a parameter-efficient deep learning framework that adapts the Segment Anything Model (SAM), a large-scale vision model, to volumetric mp-MRI data. GBT-SAM reduces input complexity by selecting fewer than 2.6\% of slices per scan while incorporating all four MRI modalities, preserving essential tumor-related information with minimal cost. Furthermore, our model is trained by a two-step fine-tuning strategy that incorporates a depth-aware module to capture inter-slice correlations and lightweight adaptation layers, resulting in just 6.5M trainable parameters, which is the lowest among SAM-based approaches. GBT-SAM achieves a Dice Score of 93.54 on the BraTS Adult Glioma dataset and demonstrates robust performance on Meningioma, Pediatric Glioma, and Sub-Saharan Glioma datasets. These results highlight GBT-SAM's potential as a computationally efficient and domain-robust framework for brain tumor segmentation using mp-MRI. Our code and models are available at https://github.com/vpulab/med-sam-brain .

Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós et al.

Scene recognition is currently one of the top-challenging research fields in computer vision. This may be due to the ambiguity between classes: images of several scene classes may share similar objects, which causes confusion among them. The problem is aggravated when images of a particular scene class are notably different. Convolutional Neural Networks (CNNs) have significantly boosted performance in scene recognition, albeit it is still far below from other recognition tasks (e.g., object or image recognition). In this paper, we describe a novel approach for scene recognition based on an end-to-end multi-modal CNN that combines image and context information by means of an attention module. Context information, in the shape of semantic segmentation, is used to gate features extracted from the RGB image by leveraging on information encoded in the semantic representation: the set of scene objects and stuff, and their relative locations. This gating process reinforces the learning of indicative scene content and enhances scene disambiguation by refocusing the receptive fields of the CNN towards them. Experimental results on four publicly available datasets show that the proposed approach outperforms every other state-of-the-art method while significantly reducing the number of network parameters. All the code and data used along this paper is available at https://github.com/vpulab/Semantic-Aware-Scene-Recognition

Roberto Alcover-Couso, Marcos Escudero-Viñolo, Juan C. SanMiguel et al.

Segmentation models are typically constrained by the categories defined during training. To address this, researchers have explored two independent approaches: adapting Vision-Language Models (VLMs) and leveraging synthetic data. However, VLMs often struggle with granularity, failing to disentangle fine-grained concepts, while synthetic data-based methods remain limited by the scope of available datasets. This paper proposes enhancing segmentation accuracy across diverse domains by integrating Vision-Language reasoning with key strategies for Unsupervised Domain Adaptation (UDA). First, we improve the fine-grained segmentation capabilities of VLMs through multi-scale contextual data, robust text embeddings with prompt augmentation, and layer-wise fine-tuning in our proposed Foundational-Retaining Open Vocabulary Semantic Segmentation (FROVSS) framework. Next, we incorporate these enhancements into a UDA framework by employing distillation to stabilize training and cross-domain mixed sampling to boost adaptability without compromising generalization. The resulting UDA-FROVSS framework is the first UDA approach to effectively adapt across domains without requiring shared categories.

Javier Montalvo, Roberto Alcover-Couso, Pablo Carballeira et al.

This paper introduces a novel synthetic dataset that captures urban scenes under a variety of weather conditions, providing pixel-perfect, ground-truth-aligned images to facilitate effective feature alignment across domains. Additionally, we propose a method for domain adaptation and generalization that takes advantage of the multiple versions of each scene, enforcing feature consistency across different weather scenarios. Our experimental results demonstrate the impact of our dataset in improving performance across several alignment metrics, addressing key challenges in domain adaptation and generalization for segmentation tasks. This research also explores critical aspects of synthetic data generation, such as optimizing the balance between the volume and variability of generated images to enhance segmentation performance. Ultimately, this work sets forth a new paradigm for synthetic data generation and domain adaptation.

Kirill Sirotkin, Marcos Escudero-Viñolo, Pablo Carballeira et al.

Foundation models trained on web-scraped datasets propagate societal biases to downstream tasks. While counterfactual generation enables bias analysis, existing methods introduce artifacts by modifying contextual elements like clothing and background. We present a localized counterfactual generation method that preserves image context by constraining counterfactual modifications to specific attribute-relevant regions through automated masking and guided inpainting. When applied to the Conceptual Captions dataset for creating gender counterfactuals, our method results in higher visual and semantic fidelity than state-of-the-art alternatives, while maintaining the performance of models trained using only real data on non-human-centric tasks. Models fine-tuned with our counterfactuals demonstrate measurable bias reduction across multiple metrics, including a decrease in gender classification disparity and balanced person preference scores, while preserving ImageNet zero-shot performance. The results establish a framework for creating balanced datasets that enable both accurate bias profiling and effective mitigation.

Kirill Sirotkin, Marcos Escudero-Viñolo, Pablo Carballeira et al.

State-of-the-art deep learning approaches for skin lesion recognition often require pretraining on larger and more varied datasets, to overcome the generalization limitations derived from the reduced size of the skin lesion imaging datasets. ImageNet is often used as the pretraining dataset, but its transferring potential is hindered by the domain gap between the source dataset and the target dermatoscopic scenario. In this work, we introduce a novel pretraining approach that sequentially trains a series of Self-Supervised Learning pretext tasks and only requires the unlabeled skin lesion imaging data. We present a simple methodology to establish an ordering that defines a pretext task curriculum. For the multi-class skin lesion classification problem, and ISIC-2019 dataset, we provide experimental evidence showing that: i) a model pretrained by a curriculum of pretext tasks outperforms models pretrained by individual pretext tasks, and ii) a model pretrained by the optimal pretext task curriculum outperforms a model pretrained on ImageNet. We demonstrate that this performance gain is related to the fact that the curriculum of pretext tasks better focuses the attention of the final model on the skin lesion. Beyond performance improvement, this strategy allows for a large reduction in the training time with respect to ImageNet pretraining, which is especially advantageous for network architectures tailored for a specific problem.

Elena Luna, Juan C. SanMiguel, Jose M. Martínez et al.

Multi-Target Multi-Camera (MTMC) vehicle tracking is an essential task of visual traffic monitoring, one of the main research fields of Intelligent Transportation Systems. Several offline approaches have been proposed to address this task; however, they are not compatible with real-world applications due to their high latency and post-processing requirements. In this paper, we present a new low-latency online approach for MTMC tracking in scenarios with partially overlapping fields of view (FOVs), such as road intersections. Firstly, the proposed approach detects vehicles at each camera. Then, the detections are merged between cameras by applying cross-camera clustering based on appearance and location. Lastly, the clusters containing different detections of the same vehicle are temporally associated to compute the tracks on a frame-by-frame basis. The experiments show promising low-latency results while addressing real-world challenges such as the a priori unknown and time-varying number of targets and the continuous state estimation of them without performing any post-processing of the trajectories.

Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós

Action recognition is currently one of the top-challenging research fields in computer vision. Convolutional Neural Networks (CNNs) have significantly boosted its performance but rely on fixed-size spatio-temporal windows of analysis, reducing CNNs temporal receptive fields. Among action recognition datasets, egocentric recorded sequences have become of important relevance while entailing an additional challenge: ego-motion is unavoidably transferred to these sequences. The proposed method aims to cope with it by estimating this ego-motion or camera motion. The estimation is used to temporally partition video sequences into motion-compensated temporal \textit{chunks} showing the action under stable backgrounds and allowing for a content-driven temporal sampling. A CNN trained in an end-to-end fashion is used to extract temporal features from each \textit{chunk}, which are late fused. This process leads to the extraction of features from the whole temporal range of an action, increasing the temporal receptive field of the network.

Andrija Gajic, Ester Gonzalez-Sosa, Diego Gonzalez-Morin et al.

The objective of this work is to segment human body parts from egocentric video using semantic segmentation networks. Our contribution is two-fold: i) we create a semi-synthetic dataset composed of more than 15, 000 realistic images and associated pixel-wise labels of egocentric human body parts, such as arms or legs including different demographic factors; ii) building upon the ThunderNet architecture, we implement a deep learning semantic segmentation algorithm that is able to perform beyond real-time requirements (16 ms for 720 x 720 images). It is believed that this method will enhance sense of presence of Virtual Environments and will constitute a more realistic solution to the standard virtual avatars.

Alejandro López-Cifuentes, Marcos Escudero-Viñolo, Jesús Bescós et al.

In the current worldwide situation, pedestrian detection has reemerged as a pivotal tool for intelligent video-based systems aiming to solve tasks such as pedestrian tracking, social distancing monitoring or pedestrian mass counting. Pedestrian detection methods, even the top performing ones, are highly sensitive to occlusions among pedestrians, which dramatically degrades their performance in crowded scenarios. The generalization of multi-camera set-ups permits to better confront occlusions by combining information from different viewpoints. In this paper, we present a multi-camera approach to globally combine pedestrian detections leveraging automatically extracted scene context. Contrarily to the majority of the methods of the state-of-the-art, the proposed approach is scene-agnostic, not requiring a tailored adaptation to the target scenario\textemdash e.g., via fine-tunning. This noteworthy attribute does not require \textit{ad hoc} training with labelled data, expediting the deployment of the proposed method in real-world situations. Context information, obtained via semantic segmentation, is used 1) to automatically generate a common Area of Interest for the scene and all the cameras, avoiding the usual need of manually defining it; and 2) to obtain detections for each camera by solving a global optimization problem that maximizes coherence of detections both in each 2D image and in the 3D scene. This process yields tightly-fitted bounding boxes that circumvent occlusions or miss-detections. Experimental results on five publicly available datasets show that the proposed approach outperforms state-of-the-art multi-camera pedestrian detectors, even some specifically trained on the target scenario, signifying the versatility and robustness of the proposed method without requiring ad-hoc annotations nor human-guided configuration.