Despite the significant progress in deep learning for dense visual recognition problems, such as semantic segmentation, traditional methods are constrained by fixed class sets. Meanwhile, vision-language foundation models, such as CLIP, have showcased remarkable effectiveness in numerous zero-shot image-level tasks, owing to their robust generalizability. Recently, a body of work has investigated utilizing these models in open-vocabulary semantic segmentation (OVSS). However, existing approaches often rely on impractical supervised pre-training or access to additional pre-trained networks. In this work, we propose a strong baseline for training-free OVSS, termed Neighbour-Aware CLIP (NACLIP), representing a straightforward adaptation of CLIP tailored for this scenario. Our method enforces localization of patches in the self-attention of CLIP's vision transformer which, despite being crucial for dense prediction tasks, has been overlooked in the OVSS literature. By incorporating design choices favouring segmentation, our approach significantly improves performance without requiring additional data, auxiliary pre-trained networks, or extensive hyperparameter tuning, making it highly practical for real-world applications. Experiments are performed on 8 popular semantic segmentation benchmarks, yielding state-of-the-art performance on most scenarios. Our code is publicly available at https://github.com/sinahmr/NACLIP .
In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin LP++, step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, LP++ operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: \url{https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git}.
State-of-the-art semi-supervised learning (SSL) approaches rely on highly confident predictions to serve as pseudo-labels that guide the training on unlabeled samples. An inherent drawback of this strategy stems from the quality of the uncertainty estimates, as pseudo-labels are filtered only based on their degree of uncertainty, regardless of the correctness of their predictions. Thus, assessing and enhancing the uncertainty of network predictions is of paramount importance in the pseudo-labeling process. In this work, we empirically demonstrate that SSL methods based on pseudo-labels are significantly miscalibrated, and formally demonstrate the minimization of the min-entropy, a lower bound of the Shannon entropy, as a potential cause for miscalibration. To alleviate this issue, we integrate a simple penalty term, which enforces the logit distances of the predictions on unlabeled samples to remain low, preventing the network predictions to become overconfident. Comprehensive experiments on a variety of SSL image classification benchmarks demonstrate that the proposed solution systematically improves the calibration performance of relevant SSL models, while also enhancing their discriminative power, being an appealing addition to tackle SSL tasks.
In this work, we present a novel approach to calibrate segmentation networks that considers the inherent challenges posed by different categories and object regions. In particular, we present a formulation that integrates class and region-wise constraints into the learning objective, with multiple penalty weights to account for class and region differences. Finding the optimal penalty weights manually, however, might be unfeasible, and potentially hinder the optimization process. To overcome this limitation, we propose an approach based on Class and Region-Adaptive constraints (CRaC), which allows to learn the class and region-wise penalty weights during training. CRaC is based on a general Augmented Lagrangian method, a well-established technique in constrained optimization. Experimental results on two popular segmentation benchmarks, and two well-known segmentation networks, demonstrate the superiority of CRaC compared to existing approaches. The code is available at: https://github.com/Bala93/CRac/
Using deep learning models pre-trained on Imagenet is the traditional solution for medical image classification to deal with data scarcity. Nevertheless, relevant literature supports that this strategy may offer limited gains due to the high dissimilarity between domains. Currently, the paradigm of adapting domain-specialized foundation models is proving to be a promising alternative. However, how to perform such knowledge transfer, and the benefits and limitations it presents, are under study. The CGI-HRDC challenge for Hypertensive Retinopathy diagnosis on fundus images introduces an appealing opportunity to evaluate the transferability of a recently released vision-language foundation model of the retina, FLAIR. In this work, we explore the potential of using FLAIR features as starting point for fundus image classification, and we compare its performance with regard to Imagenet initialization on two popular transfer learning methods: Linear Probing (LP) and Fine-Tuning (FP). Our empirical observations suggest that, in any case, the use of the traditional strategy provides performance gains. In contrast, direct transferability from FLAIR model allows gains of 2.5%. When fine-tuning the whole network, the performance gap increases up to 4%. In this case, we show that avoiding feature deterioration via LP initialization of the classifier allows the best re-use of the rich pre-trained features. Although direct transferability using LP still offers limited performance, we believe that foundation models such as FLAIR will drive the evolution of deep-learning-based fundus image analysis.
Ensuring reliable confidence scores from deep neural networks is of paramount significance in critical decision-making systems, particularly in real-world domains such as healthcare. Recent literature on calibrating deep segmentation networks has resulted in substantial progress. Nevertheless, these approaches are strongly inspired by the advancements in classification tasks, and thus their uncertainty is usually modeled by leveraging the information of individual pixels, disregarding the local structure of the object of interest. Indeed, only the recent Spatially Varying Label Smoothing (SVLS) approach considers pixel spatial relationships across classes, by softening the pixel label assignments with a discrete spatial Gaussian kernel. In this work, we first present a constrained optimization perspective of SVLS and demonstrate that it enforces an implicit constraint on soft class proportions of surrounding pixels. Furthermore, our analysis shows that SVLS lacks a mechanism to balance the contribution of the constraint with the primary objective, potentially hindering the optimization process. Based on these observations, we propose NACL (Neighbor Aware CaLibration), a principled and simple solution based on equality constraints on the logit values, which enables to control explicitly both the enforced constraint and the weight of the penalty, offering more flexibility. Comprehensive experiments on a wide variety of well-known segmentation benchmarks demonstrate the superior calibration performance of the proposed approach, without affecting its discriminative power. Furthermore, ablation studies empirically show the model agnostic nature of our approach, which can be used to train a wide span of deep segmentation networks.
Efficient transfer learning (ETL) is receiving increasing attention to adapt large pre-trained language-vision models on downstream tasks with a few labeled samples. While significant progress has been made, we reveal that state-of-the-art ETL approaches exhibit strong performance only in narrowly-defined experimental setups, and with a careful adjustment of hyperparameters based on a large corpus of labeled samples. In particular, we make two interesting, and surprising empirical observations. First, to outperform a simple Linear Probing baseline, these methods require to optimize their hyper-parameters on each target task. And second, they typically underperform -- sometimes dramatically -- standard zero-shot predictions in the presence of distributional drifts. Motivated by the unrealistic assumptions made in the existing literature, i.e., access to a large validation set and case-specific grid-search for optimal hyperparameters, we propose a novel approach that meets the requirements of real-world scenarios. More concretely, we introduce a CLass-Adaptive linear Probe (CLAP) objective, whose balancing term is optimized via an adaptation of the general Augmented Lagrangian method tailored to this context. We comprehensively evaluate CLAP on a broad span of datasets and scenarios, demonstrating that it consistently outperforms SoTA approaches, while yet being a much more efficient alternative.
Semi-supervised learning relaxes the need of large pixel-wise labeled datasets for image segmentation by leveraging unlabeled data. A prominent way to exploit unlabeled data is to regularize model predictions. Since the predictions of unlabeled data can be unreliable, uncertainty-aware schemes are typically employed to gradually learn from meaningful and reliable predictions. Uncertainty estimation methods, however, rely on multiple inferences from the model predictions that must be computed for each training step, which is computationally expensive. Moreover, these uncertainty maps capture pixel-wise disparities and do not consider global information. This work proposes a novel method to estimate segmentation uncertainty by leveraging global information from the segmentation masks. More precisely, an anatomically-aware representation is first learnt to model the available segmentation masks. The learnt representation thereupon maps the prediction of a new segmentation into an anatomically-plausible segmentation. The deviation from the plausible segmentation aids in estimating the underlying pixel-level uncertainty in order to further guide the segmentation network. The proposed method consequently estimates the uncertainty using a single inference from our representation, thereby reducing the total computation. We evaluate our method on two publicly available segmentation datasets of left atria in cardiac MRIs and of multiple organs in abdominal CTs. Our anatomically-aware method improves the segmentation accuracy over the state-of-the-art semi-supervised methods in terms of two commonly used evaluation metrics.
Foundation vision-language models are currently transforming computer vision, and are on the rise in medical imaging fueled by their very promising generalization capabilities. However, the initial attempts to transfer this new paradigm to medical imaging have shown less impressive performances than those observed in other domains, due to the significant domain shift and the complex, expert domain knowledge inherent to medical-imaging tasks. Motivated by the need for domain-expert foundation models, we present FLAIR, a pre-trained vision-language model for universal retinal fundus image understanding. To this end, we compiled 37 open-access, mostly categorical fundus imaging datasets from various sources, with up to 97 different target conditions and 284,660 images. We integrate the expert's domain knowledge in the form of descriptive textual prompts, during both pre-training and zero-shot inference, enhancing the less-informative categorical supervision of the data. Such a textual expert's knowledge, which we compiled from the relevant clinical literature and community standards, describes the fine-grained features of the pathologies as well as the hierarchies and dependencies between them. We report comprehensive evaluations, which illustrate the benefit of integrating expert knowledge and the strong generalization capabilities of FLAIR under difficult scenarios with domain shifts or unseen categories. When adapted with a lightweight linear probe, FLAIR outperforms fully-trained, dataset-focused models, more so in the few-shot regimes. Interestingly, FLAIR outperforms by a large margin more generalist, larger-scale image-language models, which emphasizes the potential of embedding experts' domain knowledge and the limitations of generalist models in medical imaging.
Despite the recent progress in incremental learning, addressing catastrophic forgetting under distributional drift is still an open and important problem. Indeed, while state-of-the-art domain incremental learning (DIL) methods perform satisfactorily within known domains, their performance largely degrades in the presence of novel domains. This limitation hampers their generalizability, and restricts their scalability to more realistic settings where train and test data are drawn from different distributions. To address these limitations, we present a novel DIL approach based on a mixture of prompt-tuned CLIP models (MoP-CLIP), which generalizes the paradigm of S-Prompting to handle both in-distribution and out-of-distribution data at inference. In particular, at the training stage we model the features distribution of every class in each domain, learning individual text and visual prompts to adapt to a given domain. At inference, the learned distributions allow us to identify whether a given test sample belongs to a known domain, selecting the correct prompt for the classification task, or from an unseen domain, leveraging a mixture of the prompt-tuned CLIP models. Our empirical evaluation reveals the poor performance of existing DIL methods under domain shift, and suggests that the proposed MoP-CLIP performs competitively in the standard DIL settings while outperforming state-of-the-art methods in OOD scenarios. These results demonstrate the superiority of MoP-CLIP, offering a robust and general solution to the problem of domain incremental learning.