Image classification models deployed in the real world may receive inputs outside the intended data distribution. For critical applications such as clinical decision making, it is important that a model can detect such out-of-distribution (OOD) inputs and express its uncertainty. In this work, we assess the capability of various state-of-the-art approaches for confidence-based OOD detection through a comparative study and in-depth analysis. First, we leverage a computer vision benchmark to reproduce and compare multiple OOD detection methods. We then evaluate their capabilities on the challenging task of disease classification using chest X-rays. Our study shows that high performance in a computer vision task does not directly translate to accuracy in a medical imaging task. We analyse factors that affect performance of the methods between the two tasks. Our results provide useful insights for developing the next generation of OOD detection methods.
Many contrastive representation learning methods learn a single global representation of an entire image. However, dense contrastive representation learning methods such as DenseCL (Wang et al., 2021) can learn better representations for tasks requiring stronger spatial localization of features, such as multi-label classification, detection, and segmentation. In this work, we study how to improve the quality of the representations learned by DenseCL by modifying the training scheme and objective function, and propose DenseCL++. We also conduct several ablation studies to better understand the effects of: (i) various techniques to form dense negative pairs among augmentations of different images, (ii) cross-view dense negative and positive pairs, and (iii) an auxiliary reconstruction task. Our results show 3.5% and 4% mAP improvement over SimCLR (Chen et al., 2020a) andDenseCL in COCO multi-label classification. In COCO and VOC segmentation tasks, we achieve 1.8% and 0.7% mIoU improvements over SimCLR, respectively.
Distance-based classification is frequently used in transductive few-shot learning (FSL). However, due to the high-dimensionality of image representations, FSL classifiers are prone to suffer from the hubness problem, where a few points (hubs) occur frequently in multiple nearest neighbour lists of other points. Hubness negatively impacts distance-based classification when hubs from one class appear often among the nearest neighbors of points from another class, degrading the classifier's performance. To address the hubness problem in FSL, we first prove that hubness can be eliminated by distributing representations uniformly on the hypersphere. We then propose two new approaches to embed representations on the hypersphere, which we prove optimize a tradeoff between uniformity and local similarity preservation -- reducing hubness while retaining class structure. Our experiments show that the proposed methods reduce hubness, and significantly improves transductive FSL accuracy for a wide range of classifiers.
The problem of artifacts in whole slide image acquisition, prevalent in both clinical workflows and research-oriented settings, necessitates human intervention and re-scanning. Overcoming this challenge requires developing quality control algorithms, that are hindered by the limited availability of relevant annotated data in histopathology. The manual annotation of ground-truth for artifact detection methods is expensive and time-consuming. This work addresses the issue by proposing a method dedicated to augmenting whole slide images with artifacts. The tool seamlessly generates and blends artifacts from an external library to a given histopathology dataset. The augmented datasets are then utilized to train artifact classification methods. The evaluation shows their usefulness in classification of the artifacts, where they show an improvement from 0.10 to 0.01 AUROC depending on the artifact type. The framework, model, weights, and ground-truth annotations are freely released to facilitate open science and reproducible research.
Large-scale vision models have become integral in many applications due to their unprecedented performance and versatility across downstream tasks. However, the robustness of these foundation models has primarily been explored for a single task, namely image classification. The vulnerability of other common vision tasks, such as semantic segmentation and depth estimation, remains largely unknown. We present a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks. Our attacks operate in the encoder embedding space and at the downstream task output level. In both cases, current state-of-the-art adversarial fine-tuning techniques tested only for classification significantly degrade clean and robust performance on other tasks. Since the purpose of a foundation model is to cater to multiple applications at once, our findings reveal the need to enhance encoder robustness more broadly. Our code is available at ${github.com/layer6ai-labs/ssl-robustness}$.
Given the severe challenges confronting the global growth security of economic crops, precise identification and prevention of plant diseases has emerged as a critical issue in artificial intelligence-enabled agricultural technology. To address the technical challenges in plant disease recognition, including small-sample learning, leaf occlusion, illumination variations, and high inter-class similarity, this study innovatively proposes a Dynamic Dual-Stream Fusion Network (DS_FusionNet). The network integrates a dual-backbone architecture, deformable dynamic fusion modules, and bidirectional knowledge distillation strategy, significantly enhancing recognition accuracy. Experimental results demonstrate that DS_FusionNet achieves classification accuracies exceeding 90% using only 10% of the PlantDisease and CIFAR-10 datasets, while maintaining 85% accuracy on the complex PlantWild dataset, exhibiting exceptional generalization capabilities. This research not only provides novel technical insights for fine-grained image classification but also establishes a robust foundation for precise identification and management of agricultural diseases.
The go-to strategy to apply deep networks in settings where uncertainty informs decisions--ensembling multiple training runs with random initializations--is ill-suited for the extremely large-scale models and practical fine-tuning workflows of today. We introduce a new cost-effective strategy for improving the uncertainty quantification and downstream decisions of a large model (e.g. a fine-tuned ViT-B): coupling it with a less accurate but much smaller "sidekick" (e.g. a fine-tuned ResNet-34) with a fraction of the computational cost. We propose aggregating the predictions of this \emph{Asymmetric Duo} by simple learned weighted averaging. Surprisingly, despite their inherent asymmetry, the sidekick model almost never harms the performance of the larger model. In fact, across five image classification benchmarks and a variety of model architectures and training schemes (including soups), Asymmetric Duos significantly improve accuracy, uncertainty quantification, and selective classification metrics with only ${\sim}10-20\%$ more computation.
We propose a general framework for unsupervised domain adaptation, which allows deep neural networks trained on a source domain to be tested on a different target domain without requiring any training annotations in the target domain. This is achieved by adding extra networks and losses that help regularize the features extracted by the backbone encoder network. To this end we propose the novel use of the recently proposed unpaired image-toimage translation framework to constrain the features extracted by the encoder network. Specifically, we require that the features extracted are able to reconstruct the images in both domains. In addition we require that the distribution of features extracted from images in the two domains are indistinguishable. Many recent works can be seen as specific cases of our general framework. We apply our method for domain adaptation between MNIST, USPS, and SVHN datasets, and Amazon, Webcam and DSLR Office datasets in classification tasks, and also between GTA5 and Cityscapes datasets for a segmentation task. We demonstrate state of the art performance on each of these datasets.
Objects and their relationships are critical contents for image understanding. A scene graph provides a structured description that captures these properties of an image. However, reasoning about the relationships between objects is very challenging and only a few recent works have attempted to solve the problem of generating a scene graph from an image. In this paper, we present a method that improves scene graph generation by explicitly modeling inter-dependency among the entire object instances. We design a simple and effective relational embedding module that enables our model to jointly represent connections among all related objects, rather than focus on an object in isolation. Our method significantly benefits the main part of the scene graph generation task: relationship classification. Using it on top of a basic Faster R-CNN, our model achieves state-of-the-art results on the Visual Genome benchmark. We further push the performance by introducing global context encoding module and geometrical layout encoding module. We validate our final model, LinkNet, through extensive ablation studies, demonstrating its efficacy in scene graph generation.
Recent studies on multi-domain facial image translation have achieved impressive results. The existing methods generally provide a discriminator with an auxiliary classifier to impose domain translation. However, these methods neglect important information regarding domain distribution matching. To solve this problem, we propose a switch generative adversarial network (SwitchGAN) with a more adaptive discriminator structure and a matched generator to perform delicate image translation among multiple domains. A feature-switching operation is proposed to achieve feature selection and fusion in our conditional modules. We demonstrate the effectiveness of our model. Furthermore, we also introduce a new capability of our generator that represents attribute intensity control and extracts content information without tailored training. Experiments on the Morph, RaFD and CelebA databases visually and quantitatively show that our extended SwitchGAN (i.e., Gated SwitchGAN) can achieve better translation results than StarGAN, AttGAN and STGAN. The attribute classification accuracy achieved using the trained ResNet-18 model and the FID score obtained using the ImageNet pretrained Inception-v3 model also quantitatively demonstrate the superior performance of our models.
News videos require efficient content organisation and retrieval systems, but their unstructured nature poses significant challenges for automated processing. This paper presents a comprehensive comparative analysis of image, video, and audio classifiers for automated news video segmentation. This work presents the development and evaluation of multiple deep learning approaches, including ResNet, ViViT, AST, and multimodal architectures, to classify five distinct segment types: advertisements, stories, studio scenes, transitions, and visualisations. Using a custom-annotated dataset of 41 news videos comprising 1,832 scene clips, our experiments demonstrate that image-based classifiers achieve superior performance (84.34\% accuracy) compared to more complex temporal models. Notably, the ResNet architecture outperformed state-of-the-art video classifiers while requiring significantly fewer computational resources. Binary classification models achieved high accuracy for transitions (94.23\%) and advertisements (92.74\%). These findings advance the understanding of effective architectures for news video segmentation and provide practical insights for implementing automated content organisation systems in media applications. These include media archiving, personalised content delivery, and intelligent video search.
Heatmap regression has been used for landmark localization for quite a while now. Most of the methods use a very deep stack of bottleneck modules for heatmap classification stage, followed by heatmap regression to extract the keypoints. In this paper, we present a single dendritic CNN, termed as Pose Conditioned Dendritic Convolution Neural Network (PCD-CNN), where a classification network is followed by a second and modular classification network, trained in an end to end fashion to obtain accurate landmark points. Following a Bayesian formulation, we disentangle the 3D pose of a face image explicitly by conditioning the landmark estimation on pose, making it different from multi-tasking approaches. Extensive experimentation shows that conditioning on pose reduces the localization error by making it agnostic to face pose. The proposed model can be extended to yield variable number of landmark points and hence broadening its applicability to other datasets. Instead of increasing depth or width of the network, we train the CNN efficiently with Mask-Softmax Loss and hard sample mining to achieve upto $15\%$ reduction in error compared to state-of-the-art methods for extreme and medium pose face images from challenging datasets including AFLW, AFW, COFW and IBUG.
Effective detection of arrhythmia is an important task in the remote monitoring of electrocardiogram (ECG). The traditional ECG recognition depends on the judgment of the clinicians' experience, but the results suffer from the probability of human error due to the fatigue. To solve this problem, an ECG signal classification method based on the images is presented to classify ECG signals into normal and abnormal beats by using two-dimensional convolutional neural networks (2D-CNNs). First, we compare the accuracy and robustness between one-dimensional ECG signal input method and two-dimensional image input method in AlexNet network. Then, in order to alleviate the overfitting problem in two-dimensional network, we initialize AlexNet-like network with weights trained on ImageNet, to fit the training ECG images and fine-tune the model, and to further improve the accuracy and robustness of ECG classification. The performance evaluated on the MIT-BIH arrhythmia database demonstrates that the proposed method can achieve the accuracy of 98% and maintain high accuracy within SNR range from 20 dB to 35 dB. The experiment shows that the 2D-CNNs initialized with AlexNet weights performs better than one-dimensional signal method without a large-scale dataset.
Convolutional neural networks typically encode an input image into a series of intermediate features with decreasing resolutions. While this structure is suited to classification tasks, it does not perform well for tasks requiring simultaneous recognition and localization (e.g., object detection). The encoder-decoder architectures are proposed to resolve this by applying a decoder network onto a backbone model designed for classification tasks. In this paper, we argue encoder-decoder architecture is ineffective in generating strong multi-scale features because of the scale-decreased backbone. We propose SpineNet, a backbone with scale-permuted intermediate features and cross-scale connections that is learned on an object detection task by Neural Architecture Search. Using similar building blocks, SpineNet models outperform ResNet-FPN models by ~3% AP at various scales while using 10-20% fewer FLOPs. In particular, SpineNet-190 achieves 52.5% AP with a MaskR-CNN detector and achieves 52.1% AP with a RetinaNet detector on COCO for a single model without test-time augmentation, significantly outperforms prior art of detectors. SpineNet can transfer to classification tasks, achieving 5% top-1 accuracy improvement on a challenging iNaturalist fine-grained dataset. Code is at: https://github.com/tensorflow/tpu/tree/master/models/official/detection.
In multimodal assistant, where vision is also one of the input modalities, the identification of user intent becomes a challenging task as visual input can influence the outcome. Current digital assistants take spoken input and try to determine the user intent from conversational or device context. So, a dataset, which includes visual input (i.e. images or videos for the corresponding questions targeted for multimodal assistant use cases, is not readily available. The research in visual question answering (VQA) and visual question generation (VQG) is a great step forward. However, they do not capture questions that a visually-abled person would ask multimodal assistants. Moreover, many times questions do not seek information from external knowledge. In this paper, we provide a new dataset, MMIU (MultiModal Intent Understanding), that contains questions and corresponding intents provided by human annotators while looking at images. We, then, use this dataset for intent classification task in multimodal digital assistant. We also experiment with various approaches for combining vision and language features including the use of multimodal transformer for classification of image-question pairs into 14 intents. We provide the benchmark results and discuss the role of visual and text features for the intent classification task on our dataset.
Alzheimer's disease and Frontotemporal dementia are two major types of dementia. Their accurate diagnosis and differentiation is crucial for determining specific intervention and treatment. However, differential diagnosis of these two types of dementia remains difficult at the early stage of disease due to similar patterns of clinical symptoms. Therefore, the automatic classification of multiple types of dementia has an important clinical value. So far, this challenge has not been actively explored. Recent development of deep learning in the field of medical image has demonstrated high performance for various classification tasks. In this paper, we propose to take advantage of two types of biomarkers: structure grading and structure atrophy. To this end, we propose first to train a large ensemble of 3D U-Nets to locally discriminate healthy versus dementia anatomical patterns. The result of these models is an interpretable 3D grading map capable of indicating abnormal brain regions. This map can also be exploited in various classification tasks using graph convolutional neural network. Finally, we propose to combine deep grading and atrophy-based classifications to improve dementia type discrimination. The proposed framework showed competitive performance compared to state-of-the-art methods for different tasks of disease detection and differential diagnosis.
Uncertainty quantification in automated image analysis is highly desired in many applications. Typically, machine learning models in classification or segmentation are only developed to provide binary answers; however, quantifying the uncertainty of the models can play a critical role for example in active learning or machine human interaction. Uncertainty quantification is especially difficult when using deep learning-based models, which are the state-of-the-art in many imaging applications. The current uncertainty quantification approaches do not scale well in high-dimensional real-world problems. Scalable solutions often rely on classical techniques, such as dropout, during inference or training ensembles of identical models with different random seeds to obtain a posterior distribution. In this paper, we show that these approaches fail to approximate the classification probability. On the contrary, we propose a scalable and intuitive framework to calibrate ensembles of deep learning models to produce uncertainty quantification measurements that approximate the classification probability. On unseen test data, we demonstrate improved calibration, sensitivity (in two out of three cases) and precision when being compared with the standard approaches. We further motivate the usage of our method in active learning, creating pseudo-labels to learn from unlabeled images and human-machine collaboration.
Labeling large image datasets with attributes such as facial age or object type is tedious and sometimes infeasible. Supervised machine learning methods provide a highly accurate solution, but require manual labels which are often unavailable. Zero-shot models (e.g., CLIP) do not require manual labels but are not as accurate as supervised ones, particularly when the attribute is numeric. We propose a new approach, CLIPPR (CLIP with Priors), which adapts zero-shot models for regression and classification on unlabelled datasets. Our method does not use any annotated images. Instead, we assume a prior over the label distribution in the dataset. We then train an adapter network on top of CLIP under two competing objectives: i) minimal change of predictions from the original CLIP model ii) minimal distance between predicted and prior distribution of labels. Additionally, we present a novel approach for selecting prompts for Vision & Language models using a distributional prior. Our method is effective and presents a significant improvement over the original model. We demonstrate an improvement of 28% in mean absolute error on the UTK age regression task. We also present promising results for classification benchmarks, improving the classification accuracy on the ImageNet dataset by 2.83%, without using any labels.
Digital pathology based on whole slide images (WSIs) plays a key role in cancer diagnosis and clinical practice. Due to the high resolution of the WSI and the unavailability of patch-level annotations, WSI classification is usually formulated as a weakly supervised problem, which relies on multiple instance learning (MIL) based on patches of a WSI. In this paper, we aim to learn an optimal patch-level feature space by integrating prototype learning with MIL. To this end, we develop a Trainable Prototype enhanced deep MIL (TPMIL) framework for weakly supervised WSI classification. In contrast to the conventional methods which rely on a certain number of selected patches for feature space refinement, we softly cluster all the instances by allocating them to their corresponding prototypes. Additionally, our method is able to reveal the correlations between different tumor subtypes through distances between corresponding trained prototypes. More importantly, TPMIL also enables to provide a more accurate interpretability based on the distance of the instances from the trained prototypes which serves as an alternative to the conventional attention score-based interpretability. We test our method on two WSI datasets and it achieves a new SOTA. GitHub repository: https://github.com/LitaoYang-Jet/TPMIL
In this article, we present a novel approach to detect starting motions of cyclists in real world traffic scenarios based on Motion History Images (MHIs). The method uses a deep Convolutional Neural Network (CNN) with a residual network architecture (ResNet), which is commonly used in image classification and detection tasks. By combining MHIs with a ResNet classifier and performing a frame by frame classification of the MHIs, we are able to detect starting motions in image sequences. The detection is performed using a wide angle stereo camera system at an urban intersection. We compare our algorithm to an existing method to detect movement transitions of pedestrians that uses MHIs in combination with a Histograms of Oriented Gradients (HOG) like descriptor and a Support Vector Machine (SVM), which we adapted to cyclists. To train and evaluate the methods a dataset containing MHIs of 394 cyclist starting motions was created. The results show that both methods can be used to detect starting motions of cyclists. Using the SVM approach, we were able to safely detect starting motions 0.506 s on average after the bicycle starts moving with an F1-score of 97.7%. The ResNet approach achieved an F1-score of 100% at an average detection time of 0.144 s. The ResNet approach outperformed the SVM approach in both robustness against false positive detections and detection time.