Mateusz Kozinski

Wei Lin, Leonid Karlinsky, Nina Shvetsova et al. · ibm-research, mit

Large scale Vision-Language (VL) models have shown tremendous success in aligning representations between visual and text modalities. This enables remarkable progress in zero-shot recognition, image generation & editing, and many other exciting tasks. However, VL models tend to over-represent objects while paying much less attention to verbs, and require additional tuning on video data for best zero-shot action recognition performance. While previous work relied on large-scale, fully-annotated data, in this work we propose an unsupervised approach. We adapt a VL model for zero-shot and few-shot action recognition using a collection of unlabeled videos and an unpaired action dictionary. Based on that, we leverage Large Language Models and VL models to build a text bag for each unlabeled video via matching, text expansion and captioning. We use those bags in a Multiple Instance Learning setup to adapt an image-text backbone to video data. Although finetuned on unlabeled video data, our resulting models demonstrate high transferability to numerous unseen zero-shot downstream tasks, improving the base VL model performance by up to 14\%, and even comparing favorably to fully-supervised baselines in both zero-shot and few-shot video recognition transfer. The code will be released later at \url{https://github.com/wlin-at/MAXI}.

Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski et al. · ibm-research, mit

Although action recognition systems can achieve top performance when evaluated on in-distribution test points, they are vulnerable to unanticipated distribution shifts in test data. However, test-time adaptation of video action recognition models against common distribution shifts has so far not been demonstrated. We propose to address this problem with an approach tailored to spatio-temporal models that is capable of adaptation on a single video sample at a step. It consists in a feature distribution alignment technique that aligns online estimates of test set statistics towards the training statistics. We further enforce prediction consistency over temporally augmented views of the same test video sample. Evaluations on three benchmark action recognition datasets show that our proposed technique is architecture-agnostic and able to significantly boost the performance on both, the state of the art convolutional architecture TANet and the Video Swin Transformer. Our proposed method demonstrates a substantial performance gain over existing test-time adaptation approaches in both evaluations of a single distribution shift and the challenging case of random distribution shifts. Code will be available at \url{https://github.com/wlin-at/ViTTA}.

Doruk Oner, Hussein Osman, Mateusz Kozinski et al.

Many biological and medical tasks require the delineation of 3D curvilinear structures such as blood vessels and neurites from image volumes. This is typically done using neural networks trained by minimizing voxel-wise loss functions that do not capture the topological properties of these structures. As a result, the connectivity of the recovered structures is often wrong, which lessens their usefulness. In this paper, we propose to improve the 3D connectivity of our results by minimizing a sum of topology-aware losses on their 2D projections. This suffices to increase the accuracy and to reduce the annotation effort required to provide the required annotated training data. Code is available at https://github.com/doruk-oner/ConnectivityOnProjections.

Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin et al.

Test-Time-Training (TTT) is an approach to cope with out-of-distribution (OOD) data by adapting a trained model to distribution shifts occurring at test-time. We propose to perform this adaptation via Activation Matching (ActMAD): We analyze activations of the model and align activation statistics of the OOD test data to those of the training data. In contrast to existing methods, which model the distribution of entire channels in the ultimate layer of the feature extractor, we model the distribution of each feature in multiple layers across the network. This results in a more fine-grained supervision and makes ActMAD attain state of the art performance on CIFAR-100C and Imagenet-C. ActMAD is also architecture- and task-agnostic, which lets us go beyond image classification, and score 15.4% improvement over previous approaches when evaluating a KITTI-trained object detector on KITTI-Fog. Our experiments highlight that ActMAD can be applied to online adaptation in realistic scenarios, requiring little data to attain its full performance.

M. Jehanzeb Mirza, Inkyu Shin, Wei Lin et al.

Our MATE is the first Test-Time-Training (TTT) method designed for 3D data, which makes deep networks trained for point cloud classification robust to distribution shifts occurring in test data. Like existing TTT methods from the 2D image domain, MATE also leverages test data for adaptation. Its test-time objective is that of a Masked Autoencoder: a large portion of each test point cloud is removed before it is fed to the network, tasked with reconstructing the full point cloud. Once the network is updated, it is used to classify the point cloud. We test MATE on several 3D object classification datasets and show that it significantly improves robustness of deep networks to several types of corruptions commonly occurring in 3D point clouds. We show that MATE is very efficient in terms of the fraction of points it needs for the adaptation. It can effectively adapt given as few as 5% of tokens of each test sample, making it extremely lightweight. Our experiments show that MATE also achieves competitive performance by adapting sparsely on the test data, which further reduces its computational overhead, making it ideal for real-time applications.

Franz Thaler, Martin Urschler, Mateusz Kozinski et al.

We tackle the challenging problem of single-source domain generalization (DG) for medical image segmentation. To this end, we aim for training a network on one domain (e.g., CT) and directly apply it to a different domain (e.g., MR) without adapting the model and without requiring images or annotations from the new domain during training. We propose a novel method for promoting DG when training deep segmentation networks, which we call SRCSM. During training, our method diversifies the source domain through semantic-aware random convolution, where different regions of a source image are augmented differently, based on their annotation labels. At test-time, we complement the randomization of the training domain via mapping the intensity of target domain images, making them similar to source domain data. We perform a comprehensive evaluation on a variety of cross-modality and cross-center generalization settings for abdominal, whole-heart and prostate segmentation, where we outperform previous DG techniques in a vast majority of experiments. Additionally, we also investigate our method when training on whole-heart CT or MR data and testing on the diastolic and systolic phase of cine MR data captured with different scanner hardware, where we make a step towards closing the domain gap in this even more challenging setting. Overall, our evaluation shows that SRCSM can be considered a new state-of-the-art in DG for medical image segmentation and, moreover, even achieves a segmentation performance that matches the performance of the in-domain baseline in several settings.

M. Jehanzeb Mirza, Leonid Karlinsky, Wei Lin et al.

Prompt ensembling of Large Language Model (LLM) generated category-specific prompts has emerged as an effective method to enhance zero-shot recognition ability of Vision-Language Models (VLMs). To obtain these category-specific prompts, the present methods rely on hand-crafting the prompts to the LLMs for generating VLM prompts for the downstream tasks. However, this requires manually composing these task-specific prompts and still, they might not cover the diverse set of visual concepts and task-specific styles associated with the categories of interest. To effectively take humans out of the loop and completely automate the prompt generation process for zero-shot recognition, we propose Meta-Prompting for Visual Recognition (MPVR). Taking as input only minimal information about the target task, in the form of its short natural language description, and a list of associated class labels, MPVR automatically produces a diverse set of category-specific prompts resulting in a strong zero-shot classifier. MPVR generalizes effectively across various popular zero-shot image recognition benchmarks belonging to widely different domains when tested with multiple LLMs and VLMs. For example, MPVR obtains a zero-shot recognition improvement over CLIP by up to 19.8% and 18.2% (5.0% and 4.5% on average over 20 datasets) leveraging GPT and Mixtral LLMs, respectively

Jakub Micorek, Horst Possegger, Dominik Narnhofer et al.

We propose a novel approach to video anomaly detection: we treat feature vectors extracted from videos as realizations of a random variable with a fixed distribution and model this distribution with a neural network. This lets us estimate the likelihood of test videos and detect video anomalies by thresholding the likelihood estimates. We train our video anomaly detector using a modification of denoising score matching, a method that injects training data with noise to facilitate modeling its distribution. To eliminate hyperparameter selection, we model the distribution of noisy video features across a range of noise levels and introduce a regularizer that tends to align the models for different levels of noise. At test time, we combine anomaly indications at multiple noise scales with a Gaussian mixture model. Running our video anomaly detector induces minimal delays as inference requires merely extracting the features and forward-propagating them through a shallow neural network and a Gaussian mixture model. Our experiments on five popular video anomaly detection benchmarks demonstrate state-of-the-art performance, both in the object-centric and in the frame-centric setup.

Stefan Leitner, M. Jehanzeb Mirza, Wei Lin et al.

In autonomous driving scenarios, current object detection models show strong performance when tested in clear weather. However, their performance deteriorates significantly when tested in degrading weather conditions. In addition, even when adapted to perform robustly in a sequence of different weather conditions, they are often unable to perform well in all of them and suffer from catastrophic forgetting. To efficiently mitigate forgetting, we propose Domain-Incremental Learning through Activation Matching (DILAM), which employs unsupervised feature alignment to adapt only the affine parameters of a clear weather pre-trained network to different weather conditions. We propose to store these affine parameters as a memory bank for each weather condition and plug-in their weather-specific parameters during driving (i.e. test time) when the respective weather conditions are encountered. Our memory bank is extremely lightweight, since affine parameters account for less than 2% of a typical object detector. Furthermore, contrary to previous domain-incremental learning approaches, we do not require the weather label when testing and propose to automatically infer the weather condition by a majority voting linear classifier.

M. Jehanzeb Mirza, Leonid Karlinsky, Wei Lin et al.

Recently, large-scale pre-trained Vision and Language (VL) models have set a new state-of-the-art (SOTA) in zero-shot visual classification enabling open-vocabulary recognition of potentially unlimited set of categories defined as simple language prompts. However, despite these great advances, the performance of these zeroshot classifiers still falls short of the results of dedicated (closed category set) classifiers trained with supervised fine tuning. In this paper we show, for the first time, how to reduce this gap without any labels and without any paired VL data, using an unlabeled image collection and a set of texts auto-generated using a Large Language Model (LLM) describing the categories of interest and effectively substituting labeled visual instances of those categories. Using our label-free approach, we are able to attain significant performance improvements over the zero-shot performance of the base VL model and other contemporary methods and baselines on a wide variety of datasets, demonstrating absolute improvement of up to 11.7% (3.8% on average) in the label-free setting. Moreover, despite our approach being label-free, we observe 1.3% average gains over leading few-shot prompting baselines that do use 5-shot supervision.

Subeesh Vasu, Mateusz Kozinski, Leonardo Citraro et al.

Most state-of-the-art approaches to road extraction from aerial images rely on a CNN trained to label road pixels as foreground and remainder of the image as background. The CNN is usually trained by minimizing pixel-wise losses, which is less than ideal to produce binary masks that preserve the road network's global connectivity. To address this issue, we introduce an Adversarial Learning (AL) strategy tailored for our purposes. A naive one would treat the segmentation network as a generator and would feed its output along with ground-truth segmentations to a discriminator. It would then train the generator and discriminator jointly. We will show that this is not enough because it does not capture the fact that most errors are local and need to be treated as such. Instead, we use a more sophisticated discriminator that returns a label pyramid describing what portions of the road network are correct at several different scales. This discriminator and the structured labels it returns are what gives our approach its edge and we will show that it outperforms state-of-the-art ones on the challenging RoadTracer dataset.

Agata Mosinska, Mateusz Kozinski, Pascal Fua

Detection of curvilinear structures in images has long been of interest. One of the most challenging aspects of this problem is inferring the graph representation of the curvilinear network. Most existing delineation approaches first perform binary segmentation of the image and then refine it using either a set of hand-designed heuristics or a separate classifier that assigns likelihood to paths extracted from the pixel-wise prediction. In our work, we bridge the gap between segmentation and path classification by training a deep network that performs those two tasks simultaneously. We show that this approach is beneficial because it enforces consistency across the whole processing pipeline. We apply our approach on roads and neurons datasets.

Agata Mosinska, Pablo Marquez-Neila, Mateusz Kozinski et al.

Delineation of curvilinear structures is an important problem in Computer Vision with multiple practical applications. With the advent of Deep Learning, many current approaches on automatic delineation have focused on finding more powerful deep architectures, but have continued using the habitual pixel-wise losses such as binary cross-entropy. In this paper we claim that pixel-wise losses alone are unsuitable for this problem because of their inability to reflect the topological impact of mistakes in the final prediction. We propose a new loss term that is aware of the higher-order topological features of linear structures. We also introduce a refinement pipeline that iteratively applies the same model over the previous delineation to refine the predictions at each step while keeping the number of parameters and the complexity of the model constant. When combined with the standard pixel-wise loss, both our new loss term and our iterative refinement boost the quality of the predicted delineations, in some cases almost doubling the accuracy as compared to the same classifier trained with the binary cross-entropy alone. We show that our approach outperforms state-of-the-art methods on a wide range of data, from microscopy to aerial images.