Medical image classification is a critical problem for healthcare, with the potential to alleviate the workload of doctors and facilitate diagnoses of patients. However, two challenges arise when deploying deep learning models to real-world healthcare applications. First, neural models tend to learn spurious correlations instead of desired features, which could fall short when generalizing to new domains (e.g., patients with different ages). Second, these black-box models lack interpretability. When making diagnostic predictions, it is important to understand why a model makes a decision for trustworthy and safety considerations. In this paper, to address these two limitations, we propose a new paradigm to build robust and interpretable medical image classifiers with natural language concepts. Specifically, we first query clinical concepts from GPT-4, then transform latent image features into explicit concepts with a vision-language model. We systematically evaluate our method on eight medical image classification datasets to verify its effectiveness. On challenging datasets with strong confounding factors, our method can mitigate spurious correlations thus substantially outperform standard visual encoders and other baselines. Finally, we show how classification with a small number of concepts brings a level of interpretability for understanding model decisions through case studies in real medical data.
Photonic computing promises faster and more energy-efficient deep neural network (DNN) inference than traditional digital hardware. Advances in photonic computing can have profound impacts on applications such as autonomous driving and defect detection that depend on fast, accurate and energy efficient execution of image segmentation models. In this paper, we investigate image segmentation on photonic accelerators to explore: a) the types of image segmentation DNN architectures that are best suited for photonic accelerators, and b) the throughput and energy efficiency of executing the different image segmentation models on photonic accelerators, along with the trade-offs involved therein. Specifically, we demonstrate that certain segmentation models exhibit negligible loss in accuracy (compared to digital float32 models) when executed on photonic accelerators, and explore the empirical reasoning for their robustness. We also discuss techniques for recovering accuracy in the case of models that do not perform well. Further, we compare throughput (inferences-per-second) and energy consumption estimates for different image segmentation workloads on photonic accelerators. We discuss the challenges and potential optimizations that can help improve the application of photonic accelerators to such computer vision tasks.
Recent deep generative models (DGMs) such as generative adversarial networks (GANs) and diffusion probabilistic models (DPMs) have shown their impressive ability in generating high-fidelity photorealistic images. Although looking appealing to human eyes, training a model on purely synthetic images for downstream image processing tasks like image classification often results in an undesired performance drop compared to training on real data. Previous works have demonstrated that enhancing a real dataset with synthetic images from DGMs can be beneficial. However, the improvements were subjected to certain circumstances and yet were not comparable to adding the same number of real images. In this work, we propose a new taxonomy to describe factors contributing to this commonly observed phenomenon and investigate it on the popular CIFAR-10 dataset. We hypothesize that the Content Gap accounts for a large portion of the performance drop when using synthetic images from DGM and propose strategies to better utilize them in downstream tasks. Extensive experiments on multiple datasets showcase that our method outperforms baselines on downstream classification tasks both in case of training on synthetic only (Synthetic-to-Real) and training on a mix of real and synthetic data (Data Augmentation), particularly in the data-scarce scenario.
The remarkable generative capabilities of denoising diffusion models have raised new concerns regarding the authenticity of the images we see every day on the Internet. However, the vast majority of existing deepfake detection models are tested against previous generative approaches (e.g. GAN) and usually provide only a "fake" or "real" label per image. We believe a more informative output would be to augment the per-image label with a localization map indicating which regions of the input have been manipulated. To this end, we frame this task as a weakly-supervised localization problem and identify three main categories of methods (based on either explanations, local scores or attention), which we compare on an equal footing by using the Xception network as the common backbone architecture. We provide a careful analysis of all the main factors that parameterize the design space: choice of method, type of supervision, dataset and generator used in the creation of manipulated images; our study is enabled by constructing datasets in which only one of the components is varied. Our results show that weakly-supervised localization is attainable, with the best performing detection method (based on local scores) being less sensitive to the looser supervision than to the mismatch in terms of dataset or generator.
Deep learning based object detectors struggle generalizing to a new target domain bearing significant variations in object and background. Most current methods align domains by using image or instance-level adversarial feature alignment. This often suffers due to unwanted background and lacks class-specific alignment. A straightforward approach to promote class-level alignment is to use high confidence predictions on unlabeled domain as pseudo-labels. These predictions are often noisy since model is poorly calibrated under domain shift. In this paper, we propose to leverage model's predictive uncertainty to strike the right balance between adversarial feature alignment and class-level alignment. We develop a technique to quantify predictive uncertainty on class assignments and bounding-box predictions. Model predictions with low uncertainty are used to generate pseudo-labels for self-training, whereas the ones with higher uncertainty are used to generate tiles for adversarial feature alignment. This synergy between tiling around uncertain object regions and generating pseudo-labels from highly certain object regions allows capturing both image and instance-level context during the model adaptation. We report thorough ablation study to reveal the impact of different components in our approach. Results on five diverse and challenging adaptation scenarios show that our approach outperforms existing state-of-the-art methods with noticeable margins.
Flow and bridge matching are a novel class of processes which encompass diffusion models. One of the main aspect of their increased flexibility is that these models can interpolate between arbitrary data distributions i.e. they generalize beyond generative modeling and can be applied to learning stochastic (and deterministic) processes of arbitrary transfer tasks between two given distributions. In this paper, we highlight that while flow and bridge matching processes preserve the information of the marginal distributions, they do \emph{not} necessarily preserve the coupling information unless additional, stronger optimality conditions are met. This can be problematic if one aims at preserving the original empirical pairing. We show that a simple modification of the matching process recovers this coupling by augmenting the velocity field (or drift) with the information of the initial sample point. Doing so, we lose the Markovian property of the process but preserve the coupling information between distributions. We illustrate the efficiency of our augmentation in learning mixture of image translation tasks.
Several applications require the super-resolution of noisy images and the preservation of geometrical and texture features. State-of-the-art super-resolution methods do not account for noise and generally enhance the output image's artefacts (e.g., aliasing, blurring). We propose a learning-based method that accounts for the presence of noise and preserves the properties of the input image, as measured by quantitative metrics (e.g., normalised crossed correlation, normalised mean squared error, peak-signal-to-noise-ration, structural similarity feature-based similarity, universal image quality). We train our network to up-sample a low-resolution noisy image while preserving its properties. We perform our tests on the Cineca Marconi100 cluster, at the 26th position in the top500 list. The experimental results show that our method outperforms learning-based methods, has comparable results with standard methods, preserves the properties of the input image as contours, brightness, and textures, and reduces the artefacts. As average quantitative metrics, our method has a PSNR value of 23.81 on the super-resolution of Gaussian noise images with a 2X up-sampling factor. In contrast, previous work has a PSNR value of 23.09 (standard method) and 21.78 (learning-based method). Our learning-based and quality-preserving super-resolution improves the high-resolution prediction of noisy images with respect to state-of-the-art methods with different noise types and up-sampling factors.
Diffusion models are generative models with impressive text-to-image synthesis capabilities and have spurred a new wave of creative methods for classical machine learning tasks. However, the best way to harness the perceptual knowledge of these generative models for visual tasks is still an open question. Specifically, it is unclear how to use the prompting interface when applying diffusion backbones to vision tasks. We find that automatically generated captions can improve text-image alignment and significantly enhance a model's cross-attention maps, leading to better perceptual performance. Our approach improves upon the current SOTA in diffusion-based semantic segmentation on ADE20K and the current overall SOTA in depth estimation on NYUv2. Furthermore, our method generalizes to the cross-domain setting; we use model personalization and caption modifications to align our model to the target domain and find improvements over unaligned baselines. Our object detection model, trained on Pascal VOC, achieves SOTA results on Watercolor2K. Our segmentation method, trained on Cityscapes, achieves SOTA results on Dark Zurich-val and Nighttime Driving. Project page: https://www.vision.caltech.edu/tadp/
Over the years, space scientists have collected terabytes of Mars data from satellites and rovers. One important set of features identified in Mars orbital images is pitted cones, which are interpreted to be mud volcanoes believed to form in regions that were once saturated in water (i.e., a lake or ocean). Identifying pitted cones globally on Mars would be of great importance, but expert geologists are unable to sort through the massive orbital image archives to identify all examples. However, this task is well suited for computer vision. Although several computer vision datasets exist for various Mars-related tasks, there is currently no open-source dataset available for cone detection/segmentation. Furthermore, previous studies trained models using data from a single region, which limits their applicability for global detection and mapping. Motivated by this, we introduce ConeQuest, the first expert-annotated public dataset to identify cones on Mars. ConeQuest consists of >13k samples from 3 different regions of Mars. We propose two benchmark tasks using ConeQuest: (i) Spatial Generalization and (ii) Cone-size Generalization. We finetune and evaluate widely-used segmentation models on both benchmark tasks. Results indicate that cone segmentation is a challenging open problem not solved by existing segmentation models, which achieve an average IoU of 52.52% and 42.55% on in-distribution data for tasks (i) and (ii), respectively. We believe this new benchmark dataset will facilitate the development of more accurate and robust models for cone segmentation. Data and code are available at https://github.com/kerner-lab/ConeQuest.
In coded aperture snapshot spectral imaging (CASSI) systems, a core problem is to recover the 3D hyperspectral image (HSI) from the 2D measurement. Current deep unfolding networks (DUNs) for the HSI reconstruction mainly suffered from three issues. Firstly, in previous DUNs, the DNNs across different stages were unable to share the feature representations learned from different stages, leading to parameter sparsity, which in turn limited their reconstruction potential. Secondly, previous DUNs fail to estimate degradation-related parameters within a unified framework, including the degradation matrix in the data subproblem and the noise level in the prior subproblem. Consequently, either the accuracy of solving the data or the prior subproblem is compromised. Thirdly, exploiting both local and non-local priors for the HSI reconstruction is crucial, and it remains a key issue to be addressed. In this paper, we first transform the DUN into a Recurrent Neural Network (RNN) by sharing parameters across stages, which allows the DNN in each stage could learn feature representation from different stages, enhancing the representativeness of the DUN. Secondly, we incorporate the Degradation Estimation Network into the RNN (DERNN), which simultaneously estimates the degradation matrix and the noise level by residual learning with reference to the sensing matrix. Thirdly, we propose a Local and Non-Local Transformer (LNLT) to effectively exploit both local and non-local priors in HSIs. By integrating the LNLT into the DERNN for solving the prior subproblem, we propose the DERNN-LNLT, which achieves state-of-the-art performance.