AI assistants that support humans in daily life are becoming increasingly feasible, driven by the rapid advancements in multimodal language models. A key challenge lies in overcoming the generic nature of these models to deliver personalized experiences. Existing approaches to personalizing large vision language models often rely on additional training stages, which limit generality and scalability, or on engineered pipelines with external pre-trained modules, which hinder deployment efficiency. In this work, we propose an efficient personalization method that leverages the model's inherent ability to capture personalized concepts. Specifically, we extract visual tokens that predominantly represent the target concept by utilizing the model's internal attention mechanisms. These tokens serve as a memory of that specific concept, enabling the model to recall and describe it when it appears in test images. We conduct a comprehensive and unified evaluation of our approach and SOTA methods across various personalization settings including single-concept, multi-concept, and video personalization, demonstrating strong performance gains with minimal personalization overhead.

3D occupancy prediction is gaining traction in autonomous driving for its ability to jointly model environment geometry and semantics. Weakly supervised methods learn 3D representations solely from multi-view 2D labels, making them ideal for data-scarce scenarios. However, distilling 2D knowledge into 3D is challenging due to limited information and noisy pseudo labels. To tackle the above challenges, we introduce MT-Occ, a single-view self-supervised 3D occupancy prediction method that enhances 2D-to-3D distillation by leveraging pretrained multi-task features and modelling task interactions. Our approach includes three effective and flexible components: 1) an effective fusion technique utilising pretrained features of multiple relevant tasks, 2) a spatial cross-task attention module for geometric-semantic distillation and 3) a view-consistent label refinement strategy to improve 2D pseudo labels. MT-Occ achieves state-of-the-art results on autonomous driving benchmarks, outperforming prior work with +16.04

Level 5 autonomy for self-driving cars requires a robust visual perception system that can parse input images under any visual condition. However, existing semantic segmentation datasets are either dominated by images captured under normal conditions or are small in scale. To address this, we introduce ACDC, the Adverse Conditions Dataset with Correspondences for training and testing semantic segmentation methods on adverse visual conditions. ACDC consists of a large set of 4006 images which are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow. Each adverse-condition image comes with a high-quality fine pixel-level semantic annotation, a corresponding image of the same scene taken under normal conditions, and a binary mask that distinguishes between intra-image regions of clear and uncertain semantic content. Thus, ACDC supports both standard semantic segmentation and the newly introduced uncertainty-aware semantic segmentation. A detailed empirical study demonstrates the challenges that the adverse domains of ACDC pose to state-of-the-art supervised and unsupervised approaches and indicates the value of our dataset in steering future progress in the field. Our dataset and benchmark are publicly available.

Personalization of Large Vision-Language Models (LVLMs) involves customizing models to recognize specific users or object instances and to generate contextually tailored responses. Existing approaches rely on time-consuming training for each item, making them impractical for real-world deployment, as reflected in current personalization benchmarks limited to object-centric single-concept evaluations. In this paper, we present a novel training-free approach to LVLM personalization called . We introduce a comprehensive, real-world benchmark designed to rigorously evaluate various aspects of the personalization task. leverages pre-trained vision foundation models to extract distinctive features, applies retrieval-augmented generation (RAG) techniques to identify instances within visual inputs, and employs visual prompting strategies to guide model outputs. Our model-agnostic vision toolkit enables efficient and flexible multi-concept personalization across both images and videos, without any additional training. We achieve state-of-the-art results, surpassing existing training-based methods.

Vision-language models (VLMs) excel in tasks such as visual question answering and image captioning. However, VLMs are often limited by their use of pretrained image encoders, like CLIP, leading to image understanding errors that hinder overall performance. On top of that, real-world applications often require the model to be continuously adapted as new and often limited data continuously arrive. To address this, we propose LoRSU (Low-Rank Adaptation with Structured Updates), a robust and computationally efficient method for selectively updating image encoders within VLMs. LoRSU introduces structured and localized parameter updates, effectively correcting performance on previously error-prone data while preserving the model's general robustness. Our approach leverages theoretical insights to identify and update only the most critical parameters, achieving significant resource efficiency. Specifically, we demonstrate that LoRSU reduces computational overhead by over 25x compared to full VLM updates, without sacrificing performance. Experimental results on VQA tasks in the few-shot continual learning setting, validate LoRSU's scalability, efficiency, and effectiveness, making it a compelling solution for image encoder adaptation in resource-constrained environments.

Visual relocalization is a key technique to autonomous driving, robotics, and virtual/augmented reality. After decades of explorations, absolute pose regression (APR), scene coordinate regression (SCR), and hierarchical methods (HMs) have become the most popular frameworks. However, in spite of high efficiency, APRs and SCRs have limited accuracy especially in large-scale outdoor scenes; HMs are accurate but need to store a large number of 2D descriptors for matching, resulting in poor efficiency. In this paper, we propose an efficient and accurate framework, called VRS-NeRF, for visual relocalization with sparse neural radiance field. Precisely, we introduce an explicit geometric map (EGM) for 3D map representation and an implicit learning map (ILM) for sparse patches rendering. In this localization process, EGP provides priors of spare 2D points and ILM utilizes these sparse points to render patches with sparse NeRFs for matching. This allows us to discard a large number of 2D descriptors so as to reduce the map size. Moreover, rendering patches only for useful points rather than all pixels in the whole image reduces the rendering time significantly. This framework inherits the accuracy of HMs and discards their low efficiency. Experiments on 7Scenes, CambridgeLandmarks, and Aachen datasets show that our method gives much better accuracy than APRs and SCRs, and close performance to HMs but is much more efficient.

Semantic Segmentation is one of the most challenging vision tasks, usually requiring large amounts of training data with expensive pixel-level annotations. With the success of foundation models and especially vision-language models, recent works attempt to achieve zero-shot semantic segmentation while requiring either large scale training or additional image/pixel-level annotations. In this work, we build a lightweight module on top of a self-supervised pretrained vision encoder to align patch features with a pre-trained text encoder. Importantly, we generate free annotations for any semantic segmentation dataset using existing foundation models and train our alignment module cost free. We use CLIP to detect objects and SAM to generate high quality object masks. Our approach can bring language-based semantics to any pre-trained vision encoder with minimal training. Our module is lightweight, uses foundation models as a sole source of supervision and shows impressive generalization capability from little training data with no annotation.

Out-of-Distribution (OOD) detection is a crucial problem for the safe deployment of machine learning models identifying samples that fall outside of the training distribution, i.e. in-distribution data (ID). Most OOD works focus on the classification models trained with Cross Entropy (CE) and attempt to fix its inherent issues. In this work we leverage powerful representation learned with Supervised Contrastive (SupCon) training and propose a holistic approach to learn a classifier robust to OOD data. We extend SupCon loss with two additional contrast terms. The first term pushes auxiliary OOD representations away from ID representations without imposing any constraints on similarities among auxiliary data. The second term pushes OOD features far from the existing class prototypes, while pushing ID representations closer to their corresponding class prototype. When auxiliary OOD data is not available, we propose feature mixing techniques to efficiently generate pseudo-OOD features. Our solution is simple and efficient and acts as a natural extension of the closed-set supervised contrastive representation learning. We compare against different OOD detection methods on the common benchmarks and show state-of-the-art results.

Vision language models (VLMs) demonstrate impressive capabilities in visual question answering and image captioning, acting as a crucial link between visual and language models. However, existing open-source VLMs heavily rely on pretrained and frozen vision encoders (such as CLIP). Despite CLIP's robustness across diverse domains, it still exhibits non-negligible image understanding errors. These errors propagate to the VLM responses, resulting in sub-optimal performance. In our work, we propose an efficient and robust method for updating vision encoders within VLMs. Our approach selectively and locally updates encoders, leading to substantial performance improvements on data where previous mistakes occurred, while maintaining overall robustness. Furthermore, we demonstrate the effectiveness of our method during continual few-shot updates. Theoretical grounding, generality, and computational efficiency characterize our approach.

In Class-Incremental Learning (CIL) an image classification system is exposed to new classes in each learning session and must be updated incrementally. Methods approaching this problem have updated both the classification head and the feature extractor body at each session of CIL. In this work, we develop a baseline method, First Session Adaptation (FSA), that sheds light on the efficacy of existing CIL approaches, and allows us to assess the relative performance contributions from head and body adaption. FSA adapts a pre-trained neural network body only on the first learning session and fixes it thereafter; a head based on linear discriminant analysis (LDA), is then placed on top of the adapted body, allowing exact updates through CIL. FSA is replay-free i.e. it does not memorize examples from previous sessions of continual learning. To empirically motivate FSA, we first consider a diverse selection of 22 image-classification datasets, evaluating different heads and body adaptation techniques in high/low-shot offline settings. We find that the LDA head performs well and supports CIL out-of-the-box. We also find that Featurewise Layer Modulation (FiLM) adapters are highly effective in the few-shot setting, and full-body adaption in the high-shot setting. Second, we empirically investigate various CIL settings including high-shot CIL and few-shot CIL, including settings that have previously been used in the literature. We show that FSA significantly improves over the state-of-the-art in 15 of the 16 settings considered. FSA with FiLM adapters is especially performant in the few-shot setting. These results indicate that current approaches to continuous body adaptation are not working as expected. Finally, we propose a measure that can be applied to a set of unlabelled inputs which is predictive of the benefits of body adaptation.

Cross entropy loss has served as the main objective function for classification-based tasks. Widely deployed for learning neural network classifiers, it shows both effectiveness and a probabilistic interpretation. Recently, after the success of self supervised contrastive representation learning methods, supervised contrastive methods have been proposed to learn representations and have shown superior and more robust performance, compared to solely training with cross entropy loss. However, cross entropy loss is still needed to train the final classification layer. In this work, we investigate the possibility of learning both the representation and the classifier using one objective function that combines the robustness of contrastive learning and the probabilistic interpretation of cross entropy loss. First, we revisit a previously proposed contrastive-based objective function that approximates cross entropy loss and present a simple extension to learn the classifier jointly. Second, we propose a new version of the supervised contrastive training that learns jointly the parameters of the classifier and the backbone of the network. We empirically show that these proposed objective functions demonstrate state-of-the-art performance and show a significant improvement over the standard cross entropy loss with more training stability and robustness in various challenging settings.

Novelty Detection methods identify samples that are not representative of a model's training set thereby flagging misleading predictions and bringing a greater flexibility and transparency at deployment time. However, research in this area has only considered Novelty Detection in the offline setting. Recently, there has been a growing realization in the computer vision community that applications demand a more flexible framework - Continual Learning - where new batches of data representing new domains, new classes or new tasks become available at different points in time. In this setting, Novelty Detection becomes more important, interesting and challenging. This work identifies the crucial link between the two problems and investigates the Novelty Detection problem under the Continual Learning setting. We formulate the Continual Novelty Detection problem and present a benchmark, where we compare several Novelty Detection methods under different Continual Learning settings. We show that Continual Learning affects the behaviour of novelty detection algorithms, while novelty detection can pinpoint insights in the behaviour of a continual learner. We further propose baselines and discuss possible research directions. We believe that the coupling of the two problems is a promising direction to bring vision models into practice.

A.1. Automatic global instance annotation For each image I ∈ RH×W (H and W are the height and width of the image), we first utilize a building detection network based on Mask-RCNN [3] to detect all potential building instances B = {b, b, ..., b}. Candidates with low confidence (lower than 0.9) are discarded. Next, we build a map of the environment with an off-the-shelf SfM library such as colmap [12]. The 3D map provides pixelwise correspondences for keypoints extracted from different images, which tell us which 2D building instances are identical in the 3D map. We then assign each identical building instance a global ID ranging from 1 (0 indicates background). Finally, for each image I , we obtain a dense segmentation map S ∈ RH×W in which each pixel Sij contains the global label l = 0, 1, ..., N − 1 of the global building instance that it belongs to (N is the number of global instances). All images in the Aachen dataset [11] are labeled automatically, though some minor mistakes exist at the boundaries due to the noise of building instance detection and structure-from-motion. However, as the quality of images in the RobotCar-Seasons (RoboCS) dataset [7] is relatively low because of motion blur and over-exposure, we can hardly detect building instances correctly with the pretrained network. In this case, we have to annotate global instances manually. Fortunately, images in the RoboCS dataset are densely sampled from videos, providing much overlap for consecutive images. Therefore, we manually label 857 out 6954 images captured by the rear camera and train our recognition module with them. Ground-truth of unlabeled images are obtained from the prediction of our trained model. We have 452 and 692 global building inFigure 1. Visualization ground-truth global instances. Global instances in Aachen [11] dataset are annotated automatically while in RoboCS [7] they are predicted the model trained on a small subset of images in database which are manually labeled.

While unlabelled data can be largely available and even abundant, the annotation process can be quite expensive and limiting. Under the assumption that some samples are more important for a given task than others, active learning targets the problem of identifying the most informative samples that one should acquire annotations for. In this work we propose a simple sample selection criterion that moves beyond the conventional reliance on model uncertainty as proxy to leverage new labels. By first accepting the model prediction and then judging its effect on the generalization error, we can better identify wrongly predicted samples. We also present a very efficient approximation to our criterion, providing a similarity-based interpretation. In addition to evaluating our method on the standard benchmarks of active learning, we consider the challenging yet realistic imbalanced data scenario. We show state-of-the-art results, especially on the imbalanced setting, and achieve better rates at identifying wrongly predicted samples than existing active learning methods. Our method is simple, model agnostic and relies on the current model status without the need for re-training from scratch.

We present Style2NeRF, an unsupervised model for one-shot recovery of 3D pose, shape and appearance of symmetric objects. Style2NeRF contains a transcoder which disentangles 2D representations from pretrained StyleGANs, then maps them to a semantically editable 3D NeRF generator. As such, the generative NeRF inherits Style-GAN’s expressiveness and image editing properties, translating them to 3D. We make four key contributions: (i) We provide a novel model to accurately estimate an object’s 3D pose, shape and appearance without any human supervision during training; (ii) We show how to map between semantically meaningful 2D and 3D representations using a novel disentangled generative NeRF; (iii) we introduce the pose and viewpoint ambiguity problem (suffered by existing 3D GAN methods) and propose a solution improving pose estimation accuracy; (iv) Finally, via transfer learning, we show our model can be trained on real car images where the pose distribution is unknown. Style2NeRF outperforms the state-of-the-art on the CARLA cars dataset as well as a fully supervised model for the task of car pose estimation on ShapeNet-cars and a new dataset of real car images.

Hierarchical frameworks consisting of both coarse and fine localization are often used as the standard pipeline for large-scale visual localization. Despite their promising performance in simple environments, they still suffer from low efficiency and accuracy in large-scale scenes, especially under challenging conditions. In this paper, we propose an efficient and accurate large-scale localization framework based on the recognition of buildings, which are not only discriminative for coarse localization but also robust for fine localization. Specifically, we assign each building instance a global ID and perform pixel-wise recognition of these global instances in the localization process. For coarse localization, we employ an efficient reference search strategy to find candidates progressively from the local map observing recognized instances instead of the whole database. For fine localization, predicted labels are further used for instance-wise feature detection and matching, allowing our model to focus on fewer but more robust keypoints for establishing correspondences. The experiments in long-term large-scale localization datasets including Aachen and RobotCar-Seasons demonstrate that our method outperforms previous approaches consistently in terms of both efficiency and accuracy.

Autonomous driving relies on building detailed models of a vehicles surroundings, including all hazards, obstacles and other road users. At present, much of the autonomous driving literature reduces the world to a collection of parametric 3D boxes. While this framework is sufficient for many driving scenarios, other important scene details (e.g. overhanging structures, open car doors, debris, potholes etc.) are not modelled. Recently deep implicit functions have been shown to be suitable for representing fine grained details at arbitrarily high resolutions using images alone. However, they have predominantly been employed in constrained situations, such as reconstructing individual objects or small-scale scenes. In this work we explore the application of deep implicit functions to larger scenes in the context of real-world autonomous driving scenarios. In particular we focus on the challenging case where only monocular images are available at test time. While most implicit function networks rely on watertight meshes for training, these are not in general available for real world scenes. We therefore propose an alternative training scheme using Li-DAR to provide approximate ground truth occupancy supervision. We also show that incorporating priors such as pre-detected object bounding boxes can improve the quality of reconstruction. Our method is evaluated on a real-world autonomous driving dataset.

In order to robustly deploy object detectors across a wide range of scenarios, they should be adaptable to shifts in the input distribution without the need to constantly annotate new data. This has motivated research in Unsupervised Domain Adaptation (UDA) algorithms for detection. UDA methods learn to adapt from labeled source domains to unlabeled target domains, by inducing alignment between detector features from source and target domains. Yet, there is no consensus on what features to align and how to do the alignment. In our work, we propose a framework that generalizes the different components commonly used by UDA methods laying the ground for an in-depth analysis of the UDA design space. Specifically, we propose a novel UDA algorithm, ViSGA, a direct implementation of our framework, that leverages the best design choices and introduces a simple but effective method to aggregate features at instance-level based on visual similarity before inducing group alignment via adversarial training. We show that both similarity-based grouping and adversarial training allows our model to focus on coarsely aligning feature groups, without being forced to match all instances across loosely aligned domains. Finally, we examine the applicability of ViSGA to the setting where labeled data are gathered from different sources. Experiments show that not only our method outperforms previous single-source approaches on Sim2Real and Adverse Weather, but also generalizes well to the multi-source setting.

Accurate prediction of pedestrian and bicyclist paths is integral to the development of reliable autonomous vehicles in dense urban environments. The interactions between vehicle and pedestrian or bicyclist have a significant impact on the trajectories of traffic participants e.g. stopping or turning to avoid collisions. Although recent datasets and trajectory prediction approaches have fostered the development of autonomous vehicles yet the amount of vehicle-pedestrian (bicyclist) interactions modeled are sparse. In this work, we propose Euro-PVI, a dataset of pedestrian and bicyclist trajectories. In particular, our dataset caters more diverse and complex interactions in dense urban scenarios compared to the existing datasets. To address the challenges in predicting future trajectories with dense interactions, we develop a joint inference model that learns an expressive multi-modal shared latent space across agents in the urban scene. This enables our Joint-β-cVAE approach to better model the distribution of future trajectories. We achieve state of the art results on the nuScenes and Euro-PVI datasets demonstrating the importance of capturing interactions between ego-vehicle and pedestrians (bicyclists) for accurate predictions.

Monocular 3D object detection has recently shown promising results, however there remain challenging problems. One of those is the lack of invariance to different camera intrinsic parameters, which can be observed across different 3D object datasets. Little effort has been made to exploit the combination of heterogeneous 3D object datasets. In contrast to general intuition, we show that more data does not automatically guarantee a better performance, but rather, methods need to have a degree of 'camera independence' in order to benefit from large and heterogeneous training data. In this paper we propose a category-level pose estimation method based on instance segmentation, using camera independent geometric reasoning to cope with the varying camera viewpoints and intrinsics of different datasets. Every pixel of an instance predicts the object dimensions, the 3D object reference points projected in 2D image space and, optionally, the local viewing angle. Camera intrinsics are only used outside of the learned network to lift the predicted 2D reference points to 3D. We surpass camera independent methods on the challenging KITTI3D benchmark and show the key benefits compared to camera dependent methods.