Jurandy Almeida

André Sacilotti, Samuel Felipe dos Santos, Nicu Sebe et al.

Unsupervised domain adaptation (UDA) in videos is a challenging task that remains not well explored compared to image-based UDA techniques. Although vision transformers (ViT) achieve state-of-the-art performance in many computer vision tasks, their use in video UDA has been little explored. Our key idea is to use transformer layers as a feature encoder and incorporate spatial and temporal transferability relationships into the attention mechanism. A Transferable-guided Attention (TransferAttn) framework is then developed to exploit the capacity of the transformer to adapt cross-domain knowledge across different backbones. To improve the transferability of ViT, we introduce a novel and effective module, named Domain Transferable-guided Attention Block (DTAB). DTAB compels ViT to focus on the spatio-temporal transferability relationship among video frames by changing the self-attention mechanism to a transferability attention mechanism. Extensive experiments were conducted on UCF-HMDB, Kinetics-Gameplay, and Kinetics-NEC Drone datasets, with different backbones, like ResNet101, I3D, and STAM, to verify the effectiveness of TransferAttn compared with state-of-the-art approaches. Also, we demonstrate that DTAB yields performance gains when applied to other state-of-the-art transformer-based UDA methods from both video and image domains. Our code is available at https://github.com/Andre-Sacilotti/transferattn-project-code.

Mateus Roder, Jurandy Almeida, Gustavo H. de Rosa et al. · microsoft-research

In the last decade, exponential data growth supplied machine learning-based algorithms' capacity and enabled their usage in daily-life activities. Additionally, such an improvement is partially explained due to the advent of deep learning techniques, i.e., stacks of simple architectures that end up in more complex models. Although both factors produce outstanding results, they also pose drawbacks regarding the learning process as training complex models over large datasets are expensive and time-consuming. Such a problem is even more evident when dealing with video analysis. Some works have considered transfer learning or domain adaptation, i.e., approaches that map the knowledge from one domain to another, to ease the training burden, yet most of them operate over individual or small blocks of frames. This paper proposes a novel approach to map the knowledge from action recognition to event recognition using an energy-based model, denoted as Spectral Deep Belief Network. Such a model can process all frames simultaneously, carrying spatial and temporal information through the learning process. The experimental results conducted over two public video dataset, the HMDB-51 and the UCF-101, depict the effectiveness of the proposed model and its reduced computational burden when compared to traditional energy-based models, such as Restricted Boltzmann Machines and Deep Belief Networks.

Luiz H. Buris, Daniel C. G. Pedronette, Joao P. Papa et al.

Deep learning architectures have achieved promising results in different areas (e.g., medicine, agriculture, and security). However, using those powerful techniques in many real applications becomes challenging due to the large labeled collections required during training. Several works have pursued solutions to overcome it by proposing strategies that can learn more for less, e.g., weakly and semi-supervised learning approaches. As these approaches do not usually address memorization and sensitivity to adversarial examples, this paper presents three deep metric learning approaches combined with Mixup for incomplete-supervision scenarios. We show that some state-of-the-art approaches in metric learning might not work well in such scenarios. Moreover, the proposed approaches outperform most of them in different datasets.

Lucas Fernando Alvarenga e Silva, Nicu Sebe, Jurandy Almeida

Convolutional Neural Networks (CNNs) have brought revolutionary advances to many research areas due to their capacity of learning from raw data. However, when those methods are applied to non-controllable environments, many different factors can degrade the model's expected performance, such as unlabeled datasets with different levels of domain shift and category shift. Particularly, when both issues occur at the same time, we tackle this challenging setup as Open Set Domain Adaptation (OSDA) problem. In general, existing OSDA approaches focus their efforts only on aligning known classes or, if they already extract possible negative instances, use them as a new category learned with supervision during the course of training. We propose a novel way to improve OSDA approaches by extracting a high-confidence set of unknown instances and using it as a hard constraint to tighten the classification boundaries of OSDA methods. Especially, we adopt a new loss constraint evaluated in three different means, (1) directly with the pristine negative instances; (2) with randomly transformed negatives using data augmentation techniques; and (3) with synthetically generated negatives containing adversarial features. We assessed all approaches in an extensive set of experiments based on OVANet, where we could observe consistent improvements for two public benchmarks, the Office-31 and Office-Home datasets, yielding absolute gains of up to 1.3% for both Accuracy and H-Score on Office-31 and 5.8% for Accuracy and 4.7% for H-Score on Office-Home.

Samuel Felipe dos Santos, Rodrigo Berriel, Thiago Oliveira-Santos et al.

Deep learning has achieved state-of-the-art performance on several computer vision tasks and domains. Nevertheless, it still has a high computational cost and demands a significant amount of parameters. Such requirements hinder the use in resource-limited environments and demand both software and hardware optimization. Another limitation is that deep models are usually specialized into a single domain or task, requiring them to learn and store new parameters for each new one. Multi-Domain Learning (MDL) attempts to solve this problem by learning a single model that is capable of performing well in multiple domains. Nevertheless, the models are usually larger than the baseline for a single domain. This work tackles both of these problems: our objective is to prune models capable of handling multiple domains according to a user defined budget, making them more computationally affordable while keeping a similar classification performance. We achieve this by encouraging all domains to use a similar subset of filters from the baseline model, up to the amount defined by the user's budget. Then, filters that are not used by any domain are pruned from the network. The proposed approach innovates by better adapting to resource-limited devices while, to our knowledge, being the only work that is capable of handling multiple domains at test time with fewer parameters and lower computational complexity than the baseline model for a single domain.

Samuel Felipe dos Santos, Nicu Sebe, Jurandy Almeida

Convolutional neural networks (CNNs) have achieved astonishing advances over the past decade, defining state-of-the-art in several computer vision tasks. CNNs are capable of learning robust representations of the data directly from the RGB pixels. However, most image data are usually available in compressed format, from which the JPEG is the most widely used due to transmission and storage purposes demanding a preliminary decoding process that have a high computational load and memory usage. For this reason, deep learning methods capable of learning directly from the compressed domain have been gaining attention in recent years. Those methods usually extract a frequency domain representation of the image, like DCT, by a partial decoding, and then make adaptation to typical CNNs architectures to work with them. One limitation of these current works is that, in order to accommodate the frequency domain data, the modifications made to the original model increase significantly their amount of parameters and computational complexity. On one hand, the methods have faster preprocessing, since the cost of fully decoding the images is avoided, but on the other hand, the cost of passing the images though the model is increased, mitigating the possible upside of accelerating the method. In this paper, we propose a further study of the computational cost of deep models designed for the frequency domain, evaluating the cost of decoding and passing the images through the network. We also propose handcrafted and data-driven techniques for reducing the computational complexity and the number of parameters for these models in order to keep them similar to their RGB baselines, leading to efficient models with a better trade off between computational cost and accuracy.

Samuel Felipe dos Santos, Rodrigo Berriel, Thiago Oliveira-Santos et al.

Deep learning has achieved state-of-the-art performance on several computer vision tasks and domains. Nevertheless, it still has a high computational cost and demands a significant amount of parameters. Such requirements hinder the use in resource-limited environments and demand both software and hardware optimization. Another limitation is that deep models are usually specialized into a single domain or task, requiring them to learn and store new parameters for each new one. Multi-Domain Learning (MDL) attempts to solve this problem by learning a single model capable of performing well in multiple domains. Nevertheless, the models are usually larger than the baseline for a single domain. This work tackles both of these problems: our objective is to prune models capable of handling multiple domains according to a user-defined budget, making them more computationally affordable while keeping a similar classification performance. We achieve this by encouraging all domains to use a similar subset of filters from the baseline model, up to the amount defined by the user's budget. Then, filters that are not used by any domain are pruned from the network. The proposed approach innovates by better adapting to resource-limited devices while being one of the few works that handles multiple domains at test time with fewer parameters and lower computational complexity than the baseline model for a single domain.

Mina Kaviani, Jurandy Almeida, Fabio L. Verdi

The Channel Quality Indicator (CQI) plays a pivotal role in 5G networks, optimizing infrastructure dynamically to ensure high Quality of Service (QoS). Recent research has focused on improving CQI estimation in 5G networks using machine learning. In this field, the selection of the proper loss function is critical for training an accurate model. Two commonly used loss functions are Mean Squared Error (MSE) and Mean Absolute Error (MAE). Roughly speaking, MSE put more weight on outliers, MAE on the majority. Here, we argue that the Huber loss function is more suitable for CQI prediction, since it combines the benefits of both MSE and MAE. To achieve this, the Huber loss transitions smoothly between MSE and MAE, controlled by a user-defined hyperparameter called delta. However, finding the right balance between sensitivity to small errors (MAE) and robustness to outliers (MSE) by manually choosing the optimal delta is challenging. To address this issue, we propose a novel loss function, named Residual-based Adaptive Huber Loss (RAHL). In RAHL, a learnable residual is added to the delta, enabling the model to adapt based on the distribution of errors in the data. Our approach effectively balances model robustness against outliers while preserving inlier data precision. The widely recognized Long Short-Term Memory (LSTM) model is employed in conjunction with RAHL, showcasing significantly improved results compared to the aforementioned loss functions. The obtained results affirm the superiority of RAHL, offering a promising avenue for enhanced CQI prediction in 5G networks.

Alan Gomes, Anderson Gonçalves, Samuel Felipe dos Santos et al.

Plant phenology-the study of recurrent life cycle events-is essential for understanding ecosystem dynamics and their responses to climate change impacts. While Unmanned Aerial Vehicles (UAVs) and near-surface cameras enable high-resolution monitoring, identifying plant species across time remains computationally challenging. State-of-the-art approaches, specifically Multi-Temporal Convolutional Networks (CNNs), rely on rigid multi-branch architectures that scale poorly with longer time series and require large spatial context windows. In this paper, we present an extensive study on optimizing Vision Transformers (ViTs) for efficient spatio-temporal vegetation pixel classification. We conducted a comprehensive ablation study analyzing seven key design dimensions, including: (i) data normalization; (ii) spectral arrangement; (iii) boundary handling; (iv) spatial context window shape and size; (v) tokenization strategies; (vi) positional encoding; and (vii) feature aggregation strategies. Our method was evaluated on two datasets from the Brazilian Cerrado biome, Serra do Cipó (aerial imagery) and Itirapina (near-surface imagery). Experimental results demonstrate that our ViT approach offers a substantial improvement in computational efficiency while maintaining competitive classification performance. Notably, our ViT reduces Floating Point Operations (FLOPs) by an order of magnitude and maintains constant parameter complexity regardless of the time series length, whereas the CNN baseline scales linearly. Our findings confirm that ViTs are a robust, scalable solution for resource-constrained phenological monitoring systems.

Públio Elon Correa da Silva, Jurandy Almeida

Deep learning (DL) technologies can transform agriculture by improving crop health monitoring and management, thus improving food safety. In this paper, we explore the potential of edge computing for real-time classification of leaf diseases using thermal imaging. We present a thermal image dataset for plant disease classification and evaluate deep learning models, including InceptionV3, MobileNetV1, MobileNetV2, and VGG-16, on resource-constrained devices like the Raspberry Pi 4B. Using pruning and quantization-aware training, these models achieve inference times up to 1.48x faster on Edge TPU Max for VGG16, and up to 2.13x faster with precision reduction on Intel NCS2 for MobileNetV1, compared to high-end GPUs like the RTX 3090, while maintaining state-of-the-art accuracy.

Lucas Fernando Alvarenga e Silva, Samuel Felipe dos Santos, Nicu Sebe et al.

Convolutional neural networks (CNNs) can learn directly from raw data, resulting in exceptional performance across various research areas. However, factors present in non-controllable environments such as unlabeled datasets with varying levels of domain and category shift can reduce model accuracy. The Open Set Domain Adaptation (OSDA) is a challenging problem that arises when both of these issues occur together. Existing OSDA approaches in literature only align known classes or use supervised training to learn unknown classes as a single new category. In this work, we introduce a new approach to improve OSDA techniques by extracting a set of high-confidence unknown instances and using it as a hard constraint to tighten the classification boundaries. Specifically, we use a new loss constraint that is evaluated in three different ways: (1) using pristine negative instances directly; (2) using data augmentation techniques to create randomly transformed negatives; and (3) with generated synthetic negatives containing adversarial features. We analyze different strategies to improve the discriminator and the training of the Generative Adversarial Network (GAN) used to generate synthetic negatives. We conducted extensive experiments and analysis on OVANet using three widely-used public benchmarks, the Office-31, Office-Home, and VisDA datasets. We were able to achieve similar H-score to other state-of-the-art methods, while increasing the accuracy on unknown categories.

Samuel Felipe dos Santos, Tiago Agostinho de Almeida, Jurandy Almeida

Universal Domain Adaptation (UniDA) seeks to transfer knowledge from a labeled source to an unlabeled target domain without assuming any relationship between their label sets, requiring models to classify known samples while rejecting unknown ones. Advanced methods like Mutual Learning Network (MLNet) use a bank of one-vs-all classifiers adapted via Open-set Entropy Minimization (OEM). However, this strategy treats all classifiers equally, diluting the learning signal. We propose the Enhanced Mutual Learning Network (E-MLNet), which integrates a dynamic weighting strategy to OEM. By leveraging the closed-set classifier's predictions, E-MLNet focuses adaptation on the most relevant class boundaries for each target sample, sharpening the distinction between known and unknown classes. We conduct extensive experiments on four challenging benchmarks: Office-31, Office-Home, VisDA-2017, and ImageCLEF. The results demonstrate that E-MLNet achieves the highest average H-scores on VisDA and ImageCLEF and exhibits superior robustness over its predecessor. E-MLNet outperforms the strong MLNet baseline in the majority of individual adaptation tasks -- 22 out of 31 in the challenging Open-Partial DA setting and 19 out of 31 in the Open-Set DA setting -- confirming the benefits of our focused adaptation strategy.

Gabriel Bellon de Carvalho, Jurandy Almeida

Segment Anything Model (SAM), a new AI model from Meta AI released in April 2023, is an ambitious tool designed to identify and separate individual objects within a given image through semantic interpretation. The advanced capabilities of SAM are the result of its training with millions of images and masks, and a few days after its release, several researchers began testing the model on medical images to evaluate its performance in this domain. With this perspective in focus -- i.e., optimizing work in the healthcare field -- this work proposes the use of this new technology to evaluate and study chest X-ray images. The approach adopted for this work, with the aim of improving the model's performance for lung segmentation, involved a transfer learning process, specifically the fine-tuning technique. After applying this adjustment, a substantial improvement was observed in the evaluation metrics used to assess SAM's performance compared to the masks provided by the datasets. The results obtained by the model after the adjustments were satisfactory and similar to cutting-edge neural networks, such as U-Net.

Eduardo Nascimento, John Just, Jurandy Almeida et al.

In precision agriculture, detecting productive crop fields is an essential practice that allows the farmer to evaluate operating performance separately and compare different seed varieties, pesticides, and fertilizers. However, manually identifying productive fields is often a time-consuming and error-prone task. Previous studies explore different methods to detect crop fields using advanced machine learning algorithms, but they often lack good quality labeled data. In this context, we propose a high-quality dataset generated by machine operation combined with Sentinel-2 images tracked over time. As far as we know, it is the first one to overcome the lack of labeled samples by using this technique. In sequence, we apply a semi-supervised classification of unlabeled data and state-of-the-art supervised and self-supervised deep learning methods to detect productive crop fields automatically. Finally, the results demonstrate high accuracy in Positive Unlabeled learning, which perfectly fits the problem where we have high confidence in the positive samples. Best performances have been found in Triplet Loss Siamese given the existence of an accurate dataset and Contrastive Learning considering situations where we do not have a comprehensive labeled dataset available.

Lucas Fernando Alvarenga e Silva, Daniel Carlos Guimarães Pedronette, Fábio Augusto Faria et al.

Deep learning (DL) has been the primary approach used in various computer vision tasks due to its relevant results achieved on many tasks. However, on real-world scenarios with partially or no labeled data, DL methods are also prone to the well-known domain shift problem. Multi-source unsupervised domain adaptation (MSDA) aims at learning a predictor for an unlabeled domain by assigning weak knowledge from a bag of source models. However, most works conduct domain adaptation leveraging only the extracted features and reducing their domain shift from the perspective of loss function designs. In this paper, we argue that it is not sufficient to handle domain shift only based on domain-level features, but it is also essential to align such information on the feature space. Unlike previous works, we focus on the network design and propose to embed Multi-Source version of DomaIn Alignment Layers (MS-DIAL) at different levels of the predictor. These layers are designed to match the feature distributions between different domains and can be easily applied to various MSDA methods. To show the robustness of our approach, we conducted an extensive experimental evaluation considering two challenging scenarios: digit recognition and object classification. The experimental results indicated that our approach can improve state-of-the-art MSDA methods, yielding relative gains of up to +30.64% on their classification accuracies.

Samuel Felipe dos Santos, Jurandy Almeida

Most image data available are often stored in a compressed format, from which JPEG is the most widespread. To feed this data on a convolutional neural network (CNN), a preliminary decoding process is required to obtain RGB pixels, demanding a high computational load and memory usage. For this reason, the design of CNNs for processing JPEG compressed data has gained attention in recent years. In most existing works, typical CNN architectures are adapted to facilitate the learning with the DCT coefficients rather than RGB pixels. Although they are effective, their architectural changes either raise the computational costs or neglect relevant information from DCT inputs. In this paper, we examine different ways of speeding up CNNs designed for DCT inputs, exploiting learning strategies to reduce the computational complexity by taking full advantage of DCT inputs. Our experiments were conducted on the ImageNet dataset. Results show that learning how to combine all DCT inputs in a data-driven fashion is better than discarding them by hand, and its combination with a reduction of layers has proven to be effective for reducing the computational costs while retaining accuracy.

Samuel Felipe dos Santos, Nicu Sebe, Jurandy Almeida

Convolutional neural networks (CNNs) have achieved astonishing advances over the past decade, defining state-of-the-art in several computer vision tasks. CNNs are capable of learning robust representations of the data directly from the RGB pixels. However, most image data are usually available in compressed format, from which the JPEG is the most widely used due to transmission and storage purposes demanding a preliminary decoding process that have a high computational load and memory usage. For this reason, deep learning methods capable of learning directly from the compressed domain have been gaining attention in recent years. Those methods usually extract a frequency domain representation of the image, like DCT, by a partial decoding, and then make adaptation to typical CNNs architectures to work with them. One limitation of these current works is that, in order to accommodate the frequency domain data, the modifications made to the original model increase significantly their amount of parameters and computational complexity. On one hand, the methods have faster preprocessing, since the cost of fully decoding the images is avoided, but on the other hand, the cost of passing the images though the model is increased, mitigating the possible upside of accelerating the method. In this paper, we propose a further study of the computational cost of deep models designed for the frequency domain, evaluating the cost of decoding and passing the images through the network. We also propose handcrafted and data-driven techniques for reducing the computational complexity and the number of parameters for these models in order to keep them similar to their RGB baselines, leading to efficient models with a better trade off between computational cost and accuracy.

Samuel Felipe dos Santos, Jurandy Almeida

Human action recognition has become one of the most active field of research in computer vision due to its wide range of applications, like surveillance, medical, industrial environments, smart homes, among others. Recently, deep learning has been successfully used to learn powerful and interpretable features for recognizing human actions in videos. Most of the existing deep learning approaches have been designed for processing video information as RGB image sequences. For this reason, a preliminary decoding process is required, since video data are often stored in a compressed format. However, a high computational load and memory usage is demanded for decoding a video. To overcome this problem, we propose a deep neural network capable of learning straight from compressed video. Our approach was evaluated on two public benchmarks, the UCF-101 and HMDB-51 datasets, demonstrating comparable recognition performance to the state-of-the-art methods, with the advantage of running up to 2 times faster in terms of inference speed.

Jurandy Almeida, Cristiano Saltori, Paolo Rota et al.

Deep learning revolution happened thanks to the availability of a massive amount of labelled data which have contributed to the development of models with extraordinary inference capabilities. Despite the public availability of a large quantity of datasets, to address specific requirements it is often necessary to generate a new set of labelled data. Quite often, the production of labels is costly and sometimes it requires specific know-how to be fulfilled. In this work, we tackle a new problem named low-budget label query that consists in suggesting to the user a small (low budget) set of samples to be labelled, from a completely unlabelled dataset, with the final goal of maximizing the classification accuracy on that dataset. In this work we first improve an Unsupervised Domain Adaptation (UDA) method to better align source and target domains using consistency constraints, reaching the state of the art on a few UDA tasks. Finally, using the previously trained model as reference, we propose a simple yet effective selection method based on uniform sampling of the prediction consistency distribution, which is deterministic and steadily outperforms other baselines as well as competing models on a large variety of publicly available datasets.

Leonardo A. Duarte, Otávio A. B. Penatti, Jurandy Almeida

In this paper, we present the Bag-of-Attributes (BoA) model for video representation aiming at video event retrieval. The BoA model is based on a semantic feature space for representing videos, resulting in high-level video feature vectors. For creating a semantic space, i.e., the attribute space, we can train a classifier using a labeled image dataset, obtaining a classification model that can be understood as a high-level codebook. This model is used to map low-level frame vectors into high-level vectors (e.g., classifier probability scores). Then, we apply pooling operations to the frame vectors to create the final bag of attributes for the video. In the BoA representation, each dimension corresponds to one category (or attribute) of the semantic space. Other interesting properties are: compactness, flexibility regarding the classifier, and ability to encode multiple semantic concepts in a single video representation. Our experiments considered the semantic space created by state-of-the-art convolutional neural networks pre-trained on 1000 object categories of ImageNet. Such deep neural networks were used to classify each video frame and then different coding strategies were used to encode the probability distribution from the softmax layer into a frame vector. Next, different pooling strategies were used to combine frame vectors in the BoA representation for a video. Results using BoA were comparable or superior to the baselines in the task of video event retrieval using the EVVE dataset, with the advantage of providing a much more compact representation.

Leonardo A. Duarte, Otávio A. B. Penatti, Jurandy Almeida

Often, videos are composed of multiple concepts or even genres. For instance, news videos may contain sports, action, nature, etc. Therefore, encoding the distribution of such concepts/genres in a compact and effective representation is a challenging task. In this sense, we propose the Bag of Genres representation, which is based on a visual dictionary defined by a genre classifier. Each visual word corresponds to a region in the classification space. The Bag of Genres video vector contains a summary of the activations of each genre in the video content. We evaluate the proposed method for video genre retrieval using the dataset of MediaEval Tagging Task of 2012 and for video event retrieval using the EVVE dataset. Results show that the proposed method achieves results comparable or superior to state-of-the-art methods, with the advantage of providing a much more compact representation than existing features.