Shai Avidan

Itai Lang, Dror Aiger, Forrester Cole et al.

Scene flow estimation is a long-standing problem in computer vision, where the goal is to find the 3D motion of a scene from its consecutive observations. Recently, there have been efforts to compute the scene flow from 3D point clouds. A common approach is to train a regression model that consumes source and target point clouds and outputs the per-point translation vector. An alternative is to learn point matches between the point clouds concurrently with regressing a refinement of the initial correspondence flow. In both cases, the learning task is very challenging since the flow regression is done in the free 3D space, and a typical solution is to resort to a large annotated synthetic dataset. We introduce SCOOP, a new method for scene flow estimation that can be learned on a small amount of data without employing ground-truth flow supervision. In contrast to previous work, we train a pure correspondence model focused on learning point feature representation and initialize the flow as the difference between a source point and its softly corresponding target point. Then, in the run-time phase, we directly optimize a flow refinement component with a self-supervised objective, which leads to a coherent and accurate flow field between the point clouds. Experiments on widespread datasets demonstrate the performance gains achieved by our method compared to existing leading techniques while using a fraction of the training data. Our code is publicly available at https://github.com/itailang/SCOOP.

Tomer Stolik, Itai Lang, Shai Avidan

A triangular mesh is one of the most popular 3D data representations. As such, the deployment of deep neural networks for mesh processing is widely spread and is increasingly attracting more attention. However, neural networks are prone to adversarial attacks, where carefully crafted inputs impair the model's functionality. The need to explore these vulnerabilities is a fundamental factor in the future development of 3D-based applications. Recently, mesh attacks were studied on the semantic level, where classifiers are misled to produce wrong predictions. Nevertheless, mesh surfaces possess complex geometric attributes beyond their semantic meaning, and their analysis often includes the need to encode and reconstruct the geometry of the shape. We propose a novel framework for a geometric adversarial attack on a 3D mesh autoencoder. In this setting, an adversarial input mesh deceives the autoencoder by forcing it to reconstruct a different geometric shape at its output. The malicious input is produced by perturbing a clean shape in the spectral domain. Our method leverages the spectral decomposition of the mesh along with additional mesh-related properties to obtain visually credible results that consider the delicacy of surface distortions. Our code is publicly available at https://github.com/StolikTomer/SAGA.

Yakir Gorski, Amir Jevnisek, Shai Avidan

MLaaS Service Providers (SPs) holding a Neural Network would like to keep the Neural Network weights secret. On the other hand, users wish to utilize the SPs' Neural Network for inference without revealing their data. Multi-Party Computation (MPC) offers a solution to achieve this. Computations in MPC involve communication, as the parties send data back and forth. Non-linear operations are usually the main bottleneck requiring the bulk of communication bandwidth. In this paper, we focus on ResNets, which serve as the backbone for many Computer Vision tasks, and we aim to reduce their non-linear components, specifically, the number of ReLUs. Our key insight is that spatially close pixels exhibit correlated ReLU responses. Building on this insight, we replace the per-pixel ReLU operation with a ReLU operation per patch. We term this approach 'Block-ReLU'. Since different layers in a Neural Network correspond to different feature hierarchies, it makes sense to allow patch-size flexibility for the various layers of the Neural Network. We devise an algorithm to choose the optimal set of patch sizes through a novel reduction of the problem to the Knapsack Problem. We demonstrate our approach in the semi-honest secure 3-party setting for four problems: Classifying ImageNet using ResNet50 backbone, classifying CIFAR100 using ResNet18 backbone, Semantic Segmentation of ADE20K using MobileNetV2 backbone, and Semantic Segmentation of Pascal VOC 2012 using ResNet50 backbone. Our approach achieves competitive performance compared to a handful of competitors. Our source code is publicly available: https://github.com/yg320/secure_inference.

Or Hirschorn, Shai Avidan

Video anomaly detection is an ill-posed problem because it relies on many parameters such as appearance, pose, camera angle, background, and more. We distill the problem to anomaly detection of human pose, thus decreasing the risk of nuisance parameters such as appearance affecting the result. Focusing on pose alone also has the side benefit of reducing bias against distinct minority groups. Our model works directly on human pose graph sequences and is exceptionally lightweight (~1K parameters), capable of running on any machine able to run the pose estimation with negligible additional resources. We leverage the highly compact pose representation in a normalizing flows framework, which we extend to tackle the unique characteristics of spatio-temporal pose data and show its advantages in this use case. The algorithm is quite general and can handle training data of only normal examples as well as a supervised setting that consists of labeled normal and abnormal examples. We report state-of-the-art results on two anomaly detection benchmarks - the unsupervised ShanghaiTech dataset and the recent supervised UBnormal dataset.

Shimon Malnick, Matan Rusanovsky, Ohad Fried et al.

Flow matching models learn to transport samples from a simple prior distribution to a complex data distribution. When prior-data pairs are coupled via optimal transport (OT), the learned trajectories are straight and non-crossing, enabling fast, even single-step, generation. However, computing the OT coupling in high dimensions is intractable, and existing methods attempt to solve the OT problem, at the cost of persistent bias or significant overhead. Rather than solving for the OT coupling, we reformulate the problem. Once the prior is treated as a design choice rather than a fixed input, the OT coupling between prior and data is no longer unique. Many priors admit an OT-optimal identity coupling to the data, leaving us free to choose one that is also tractable to sample. We identify low-frequency projection of natural images as such a choice. The identity coupling between data and its low-frequency representation is empirically OT-optimal, the prior is structured enough to be sampled by a lightweight model at inference, and the remaining flow-matching task reduces to synthesizing high-frequency detail. Interpolating the prior with Gaussian noise further improves generation quality while preserving the OT coupling. The approach requires no modifications to the flow model itself, and integrates naturally with latent-space models, classifier-free guidance, and one-step generation frameworks. Across all benchmarks, our method reduces trajectory curvature by more than $2\times$ compared to existing flow matching methods, yielding better generation quality in the few-step regime.

Or Hirschorn, Shai Avidan

Traditional 2D pose estimation models are limited by their category-specific design, making them suitable only for predefined object categories. This restriction becomes particularly challenging when dealing with novel objects due to the lack of relevant training data. To address this limitation, category-agnostic pose estimation (CAPE) was introduced. CAPE aims to enable keypoint localization for arbitrary object categories using a few-shot single model, requiring minimal support images with annotated keypoints. We present a significant departure from conventional CAPE techniques, which treat keypoints as isolated entities, by treating the input pose data as a graph. We leverage the inherent geometrical relations between keypoints through a graph-based network to break symmetry, preserve structure, and better handle occlusions. We validate our approach on the MP-100 benchmark, a comprehensive dataset comprising over 20,000 images spanning over 100 categories. Our solution boosts performance by 0.98% under a 1-shot setting, achieving a new state-of-the-art for CAPE. Additionally, we enhance the dataset with skeleton annotations. Our code and data are publicly available.

Amir Jevnisek, Shai Avidan

The increasing popularity of facial manipulation (Deepfakes) and synthetic face creation raises the need to develop robust forgery detection solutions. Crucially, most work in this domain assume that the Deepfakes in the test set come from the same Deepfake algorithms that were used for training the network. This is not how things work in practice. Instead, we consider the case where the network is trained on one Deepfake algorithm, and tested on Deepfakes generated by another algorithm. Typically, supervised techniques follow a pipeline of visual feature extraction from a deep backbone, followed by a binary classification head. Instead, our algorithm aggregates features extracted across all layers of one backbone network to detect a fake. We evaluate our approach on two domains of interest - Deepfake detection and Synthetic image detection, and find that we achieve SOTA results.

Amnon Drory, Shai Avidan, Raja Giryes

Rigid Point Cloud Registration (PCR) algorithms aim to estimate the 6-DOF relative motion between two point clouds, which is important in various fields, including autonomous driving. Recent years have seen a significant improvement in global PCR algorithms, i.e. algorithms that can handle a large relative motion. This has been demonstrated in various scenarios, including indoor scenes, but has only been minimally tested in the Automotive setting, where point clouds are produced by vehicle-mounted LiDAR sensors. In this work, we aim to answer questions that are important for automotive applications, including: which of the new algorithms is the most accurate, and which is fastest? How transferable are deep-learning approaches, e.g. what happens when you train a network with data from Boston, and run it in a vehicle in Singapore? How small can the overlap between point clouds be before the algorithms start to deteriorate? To what extent are the algorithms rotation invariant? Our results are at times surprising. When comparing robust parameter estimation methods for registration, we find that the fastest and most accurate is not one of the newest approaches. Instead, it is a modern variant of the well known RANSAC technique. We also suggest a new outlier filtering method, Grid-Prioritized Filtering (GPF), to further improve it. An additional contribution of this work is an algorithm for selecting challenging sets of frame-pairs from automotive LiDAR datasets. This enables meaningful benchmarking in the Automotive LiDAR setting, and can also improve training for learning algorithms.

Tamir Cohen, Leo Segre, Shay Shomer-Chai et al.

Reconstructing accurate 3D models of large-scale real-world scenes from unstructured, in-the-wild imagery remains a core challenge in computer vision, especially when the input views have little or no overlap. In such cases, existing reconstruction pipelines often produce multiple disconnected partial reconstructions or erroneously merge non-overlapping regions into overlapping geometry. In this work, we propose a framework that grounds each partial reconstruction to a complete reference model of the scene, enabling globally consistent alignment even in the absence of visual overlap. We obtain reference models from dense, geospatially accurate pseudo-synthetic renderings derived from Google Earth Studio. These renderings provide full scene coverage but differ substantially in appearance from real-world photographs. Our key insight is that, despite this significant domain gap, both domains share the same underlying scene semantics. We represent the reference model using 3D Gaussian Splatting, augmenting each Gaussian with semantic features, and formulate alignment as an inverse feature-based optimization scheme that estimates a global 6DoF pose and scale while keeping the reference model fixed. Furthermore, we introduce the WikiEarth dataset, which registers existing partial 3D reconstructions with pseudo-synthetic reference models. We demonstrate that our approach consistently improves global alignment when initialized with various classical and learning-based pipelines, while mitigating failure modes of state-of-the-art end-to-end models. All code and data will be released.

Leo Segre, Or Hirschorn, Shai Avidan

Foundation models are vital tools in various Computer Vision applications. They take as input a single RGB image and output a deep feature representation that is useful for various applications. However, in case we have multiple views of the same 3D scene, they operate on each image independently and do not always produce consistent features for the same 3D point. We propose a way to convert a Foundation Model into a Multi-View Foundation Model. Such a model takes as input a set of images and outputs a feature map for each image such that the features of corresponding points are as consistent as possible. This approach bypasses the need to build a consistent 3D model of the features and allows direct manipulation in the image space. Specifically, we show how to augment Transformers-based foundation models (i.e., DINO, SAM, CLIP) with intermediate 3D-aware attention layers that help match features across different views. As leading examples, we show surface normal estimation and multi-view segmentation tasks. Quantitative experiments show that our method improves feature matching considerably compared to current foundation models.

Amit Shomer, Shai Avidan

We propose SampleDepth, a Convolutional Neural Network (CNN), that is suited for an adaptive LiDAR. Typically,LiDAR sampling strategy is pre-defined, constant and independent of the observed scene. Instead of letting a LiDAR sample the scene in this agnostic fashion, SampleDepth determines, adaptively, where it is best to sample the current frame. To do that, SampleDepth uses depth samples from previous time steps to predict a sampling mask for the current frame. Crucially, SampleDepth is trained to optimize the performance of a depth completion downstream task. SampleDepth is evaluated on two different depth completion networks and two LiDAR datasets, KITTI Depth Completion and the newly introduced synthetic dataset, SHIFT. We show that SampleDepth is effective and suitable for different depth completion downstream tasks.

Shimon Malnick, Shai Avidan, Ohad Fried

We propose an algorithm for taming Normalizing Flow models - changing the probability that the model will produce a specific image or image category. We focus on Normalizing Flows because they can calculate the exact generation probability likelihood for a given image. We demonstrate taming using models that generate human faces, a subdomain with many interesting privacy and bias considerations. Our method can be used in the context of privacy, e.g., removing a specific person from the output of a model, and also in the context of debiasing by forcing a model to output specific image categories according to a given target distribution. Taming is achieved with a fast fine-tuning process without retraining the model from scratch, achieving the goal in a matter of minutes. We evaluate our method qualitatively and quantitatively, showing that the generation quality remains intact, while the desired changes are applied.

Yonathan Bornfeld, Shai Avidan

Private Inference (PI) uses cryptographic primitives to perform privacy preserving machine learning. In this setting, the owner of the network runs inference on the data of the client without learning anything about the data and without revealing any information about the model. It has been observed that a major computational bottleneck of PI is the calculation of the gate (i.e., ReLU), so a considerable amount of effort have been devoted to reducing the number of ReLUs in a given network. We focus on the DReLU, which is the non-linear step function of the ReLU and show that one DReLU can serve many ReLU operations. We suggest a new activation module where the DReLU operation is only performed on a subset of the channels (Prototype channels), while the rest of the channels (replicate channels) replicates the DReLU of each of their neurons from the corresponding neurons in one of the prototype channels. We then extend this idea to work across different layers. We show that this formulation can drastically reduce the number of DReLU operations in resnet type network. Furthermore, our theoretical analysis shows that this new formulation can solve an extended version of the XOR problem, using just one non-linearity and two neurons, something that traditional formulations and some PI specific methods cannot achieve. We achieve new SOTA results on several classification setups, and achieve SOTA results on image segmentation.

Vlad Rakhlin, Amir Jevnisek, Shai Avidan

ReLU activations are the main bottleneck in Private Inference that is based on ResNet networks. This is because they incur significant inference latency. Reducing ReLU count is a discrete optimization problem, and there are two common ways to approach it. Most current state-of-the-art methods are based on a smooth approximation that jointly optimizes network accuracy and ReLU budget at once. However, the last hard thresholding step of the optimization usually introduces a large performance loss. We take an alternative approach that works directly in the discrete domain by leveraging Coordinate Descent as our optimization framework. In contrast to previous methods, this yields a sparse solution by design. We demonstrate, through extensive experiments, that our method is State of the Art on common benchmarks.

Dvir Samuel, Barak Meiri, Haggai Maron et al.

Diffusion inversion is the problem of taking an image and a text prompt that describes it and finding a noise latent that would generate the exact same image. Most current deterministic inversion techniques operate by approximately solving an implicit equation and may converge slowly or yield poor reconstructed images. We formulate the problem by finding the roots of an implicit equation and devlop a method to solve it efficiently. Our solution is based on Newton-Raphson (NR), a well-known technique in numerical analysis. We show that a vanilla application of NR is computationally infeasible while naively transforming it to a computationally tractable alternative tends to converge to out-of-distribution solutions, resulting in poor reconstruction and editing. We therefore derive an efficient guided formulation that fastly converges and provides high-quality reconstructions and editing. We showcase our method on real image editing with three popular open-sourced diffusion models: Stable Diffusion, SDXL-Turbo, and Flux with different deterministic schedulers. Our solution, Guided Newton-Raphson Inversion, inverts an image within 0.4 sec (on an A100 GPU) for few-step models (SDXL-Turbo and Flux.1), opening the door for interactive image editing. We further show improved results in image interpolation and generation of rare objects.

Leo Segre, Shai Avidan

Neural Radiance Fields (NeRF) have advanced photorealistic novel view synthesis, but their reliance on photometric reconstruction introduces artifacts, commonly known as "floaters". These artifacts degrade novel view quality, especially in areas unseen by the training cameras. We present a fast, post-hoc NeRF cleanup method that eliminates such artifacts by enforcing our Free Space Prior, effectively minimizing floaters without disrupting the NeRF's representation of observed regions. Unlike existing approaches that rely on either Maximum Likelihood (ML) estimation to fit the data or a complex, local data-driven prior, our method adopts a Maximum-a-Posteriori (MAP) approach, selecting the optimal model parameters under a simple global prior assumption that unseen regions should remain empty. This enables our method to clean artifacts in both seen and unseen areas, enhancing novel view quality even in challenging scene regions. Our method is comparable with existing NeRF cleanup models while being 2.5x faster in inference time, requires no additional memory beyond the original NeRF, and achieves cleanup training in less than 30 seconds. Our code will be made publically available.

Matan Rusanovsky, Shimon Malnick, Shai Avidan

Vision-language models have achieved remarkable success in cross-modal understanding. Yet, these models remain limited to object-level or region-level grounding, lacking the capability for pixel-precise keypoint comprehension through natural language. We introduce a novel framework for pixel level grounding. The framework consists of two complementary components: a Point Descriptor that generates rich, contextual descriptions of individual keypoints, and a Point Localizer that regresses precise pixel coordinates from these descriptions. Unlike prior work that relies on templated prompts or keypoint names, our approach produces free-form, coarse-to-fine descriptions that situate keypoints within their visual context. Since there is no available dataset to train such a system, we introduce LlamaPointInPart, a carefully curated dataset of 20K+ image-keypoint-description triplets synthesized from multiple vision-language models, capturing multi-scale information from scene-level context to visual features around the keypoint. For cross-category generalization, we optimize the Point Descriptor on AP-10K via GRPO, using the frozen Point Localizer as a reward model to produce descriptions that maximize localization accuracy. To evaluate our results we establish a new evaluation protocol. Instead of comparing the text description produced by our method to the ground truth, we use the localizer to determine how close is the predicted point generated to the ground truth point. Experiments demonstrate superior performance compared to baseline models on LlamaPointInPart.The bidirectional nature of our framework should enable future applications in both keypoint-guided image understanding and language-guided precise localization. Our code and dataset are publicly available at https://github.com/matanr/Talking_Points.

Yishai Lavi, Leo Segre, Shai Avidan

3D Gaussian Splatting (3D-GS) has revolutionized novel view synthesis with its efficient, explicit representation. However, it lacks frequency interpretability, making it difficult to separate low-frequency structures from fine details. We introduce a frequency-decomposed 3D-GS framework that groups 3D Gaussians that correspond to subbands in the Laplacian Pyrmaids of the input images. Our approach enforces coherence within each subband (i.e., group of 3D Gaussians) through dedicated regularization, ensuring well-separated frequency components. We extend color values to both positive and negative ranges, allowing higher-frequency layers to add or subtract residual details. To stabilize optimization, we employ a progressive training scheme that refines details in a coarse-to-fine manner. Beyond interpretability, this frequency-aware design unlocks a range of practical benefits. Explicit frequency separation enables advanced 3D editing and stylization, allowing precise manipulation of specific frequency bands. It also supports dynamic level-of-detail control for progressive rendering, streaming, foveated rendering and fast geometry interaction. Through extensive experiments, we demonstrate that our method provides improved control and flexibility for emerging applications in scene editing and interactive rendering. Our code will be made publicly available.

Or Hirschorn, Shai Avidan

Category-Agnostic Pose Estimation (CAPE) localizes keypoints across diverse object categories with a single model, using one or a few annotated support images. Recent works have shown that using a pose graph (i.e., treating keypoints as nodes in a graph rather than isolated points) helps handle occlusions and break symmetry. However, these methods assume a static pose graph with equal-weight edges, leading to suboptimal results. We introduce EdgeCape, a novel framework that overcomes these limitations by predicting the graph's edge weights which optimizes localization. To further leverage structural priors, we propose integrating Markovian Structural Bias, which modulates the self-attention interaction between nodes based on the number of hops between them. We show that this improves the model's ability to capture global spatial dependencies. Evaluated on the MP-100 benchmark, which includes 100 categories and over 20K images, EdgeCape achieves state-of-the-art results in the 1-shot setting and leads among similar-sized methods in the 5-shot setting, significantly improving keypoint localization accuracy. Our code is publicly available.

Matan Jacob Cohen, Shai Avidan

Anomaly detection is a well-established research area that seeks to identify samples outside of a predetermined distribution. An anomaly detection pipeline is comprised of two main stages: (1) feature extraction and (2) normality score assignment. Recent papers used pre-trained networks for feature extraction achieving state-of-the-art results. However, the use of pre-trained networks does not fully-utilize the normal samples that are available at train time. This paper suggests taking advantage of this information by using teacher-student training. In our setting, a pretrained teacher network is used to train a student network on the normal training samples. Since the student network is trained only on normal samples, it is expected to deviate from the teacher network in abnormal cases. This difference can serve as a complementary representation to the pre-trained feature vector. Our method -- Transformaly -- exploits a pre-trained Vision Transformer (ViT) to extract both feature vectors: the pre-trained (agnostic) features and the teacher-student (fine-tuned) features. We report state-of-the-art AUROC results in both the common unimodal setting, where one class is considered normal and the rest are considered abnormal, and the multimodal setting, where all classes but one are considered normal, and just one class is considered abnormal. The code is available at https://github.com/MatanCohen1/Transformaly.

Itai Lang, Dvir Ginzburg, Shai Avidan et al.

We present a new method for real-time non-rigid dense correspondence between point clouds based on structured shape construction. Our method, termed Deep Point Correspondence (DPC), requires a fraction of the training data compared to previous techniques and presents better generalization capabilities. Until now, two main approaches have been suggested for the dense correspondence problem. The first is a spectral-based approach that obtains great results on synthetic datasets but requires mesh connectivity of the shapes and long inference processing time while being unstable in real-world scenarios. The second is a spatial approach that uses an encoder-decoder framework to regress an ordered point cloud for the matching alignment from an irregular input. Unfortunately, the decoder brings considerable disadvantages, as it requires a large amount of training data and struggles to generalize well in cross-dataset evaluations. DPC's novelty lies in its lack of a decoder component. Instead, we use latent similarity and the input coordinates themselves to construct the point cloud and determine correspondence, replacing the coordinate regression done by the decoder. Extensive experiments show that our construction scheme leads to a performance boost in comparison to recent state-of-the-art correspondence methods. Our code is publicly available at https://github.com/dvirginz/DPC.

Itai Lang, Uriel Kotlicki, Shai Avidan

Deep neural networks are prone to adversarial examples that maliciously alter the network's outcome. Due to the increasing popularity of 3D sensors in safety-critical systems and the vast deployment of deep learning models for 3D point sets, there is a growing interest in adversarial attacks and defenses for such models. So far, the research has focused on the semantic level, namely, deep point cloud classifiers. However, point clouds are also widely used in a geometric-related form that includes encoding and reconstructing the geometry. In this work, we are the first to consider the problem of adversarial examples at a geometric level. In this setting, the question is how to craft a small change to a clean source point cloud that leads, after passing through an autoencoder model, to the reconstruction of a different target shape. Our attack is in sharp contrast to existing semantic attacks on 3D point clouds. While such works aim to modify the predicted label by a classifier, we alter the entire reconstructed geometry. Additionally, we demonstrate the robustness of our attack in the case of defense, where we show that remnant characteristics of the target shape are still present at the output after applying the defense to the adversarial input. Our code is publicly available at https://github.com/itailang/geometric_adv.

Sharon Ayzik, Shai Avidan

We present a Deep Image Compression neural network that relies on side information, which is only available to the decoder. We base our algorithm on the assumption that the image available to the encoder and the image available to the decoder are correlated, and we let the network learn these correlations in the training phase. Then, at run time, the encoder side encodes the input image without knowing anything about the decoder side image and sends it to the decoder. The decoder then uses the encoded input image and the side information image to reconstruct the original image. This problem is known as Distributed Source Coding in Information Theory, and we discuss several use cases for this technology. We compare our algorithm to several image compression algorithms and show that adding decoder-only side information does indeed improve results. Our code is publicly available at https://github.com/ayziksha/DSIN.

Itai Lang, Asaf Manor, Shai Avidan

There is a growing number of tasks that work directly on point clouds. As the size of the point cloud grows, so do the computational demands of these tasks. A possible solution is to sample the point cloud first. Classic sampling approaches, such as farthest point sampling (FPS), do not consider the downstream task. A recent work showed that learning a task-specific sampling can improve results significantly. However, the proposed technique did not deal with the non-differentiability of the sampling operation and offered a workaround instead. We introduce a novel differentiable relaxation for point cloud sampling that approximates sampled points as a mixture of points in the primary input cloud. Our approximation scheme leads to consistently good results on classification and geometry reconstruction applications. We also show that the proposed sampling method can be used as a front to a point cloud registration network. This is a challenging task since sampling must be consistent across two different point clouds for a shared downstream task. In all cases, our approach outperforms existing non-learned and learned sampling alternatives. Our code is publicly available at https://github.com/itailang/SampleNet.

Oren Dovrat, Itai Lang, Shai Avidan

Processing large point clouds is a challenging task. Therefore, the data is often sampled to a size that can be processed more easily. The question is how to sample the data? A popular sampling technique is Farthest Point Sampling (FPS). However, FPS is agnostic to a downstream application (classification, retrieval, etc.). The underlying assumption seems to be that minimizing the farthest point distance, as done by FPS, is a good proxy to other objective functions. We show that it is better to learn how to sample. To do that, we propose a deep network to simplify 3D point clouds. The network, termed S-NET, takes a point cloud and produces a smaller point cloud that is optimized for a particular task. The simplified point cloud is not guaranteed to be a subset of the original point cloud. Therefore, we match it to a subset of the original points in a post-processing step. We contrast our approach with FPS by experimenting on two standard data sets and show significantly better results for a variety of applications. Our code is publicly available at: https://github.com/orendv/learning_to_sample

Matan Rusanovsky, Shimon Malnick, Amir Jevnisek et al.

Diffusion models dominate the space of text-to-image generation, yet they may produce undesirable outputs, including explicit content or private data. To mitigate this, concept ablation techniques have been explored to limit the generation of certain concepts. In this paper, we reveal that the erased concept information persists in the model and that erased concept images can be generated using the right latent. Utilizing inversion methods, we show that there exist latent seeds capable of generating high quality images of erased concepts. Moreover, we show that these latents have likelihoods that overlap with those of images outside the erased concept. We extend this to demonstrate that for every image from the erased concept set, we can generate many seeds that generate the erased concept. Given the vast space of latents capable of generating ablated concept images, our results suggest that fully erasing concept information may be intractable, highlighting possible vulnerabilities in current concept ablation techniques.

Leo Segre, Shai Avidan

3D scene registration is a fundamental problem in computer vision that seeks the best 6-DoF alignment between two scenes. This problem was extensively investigated in the case of point clouds and meshes, but there has been relatively limited work regarding Neural Radiance Fields (NeRF). In this paper, we consider the problem of rigid registration between two NeRFs when the position of the original cameras is not given. Our key novelty is the introduction of Viewshed Fields (VF), an implicit function that determines, for each 3D point, how likely it is to be viewed by the original cameras. We demonstrate how VF can help in the various stages of NeRF registration, with an extensive evaluation showing that VF-NeRF achieves SOTA results on various datasets with different capturing approaches such as LLFF and Objaverese.

Inbal Cohen, Boaz Meivar, Peihan Tu et al.

Segment Anything Models (SAM) have achieved remarkable success in object segmentation tasks across diverse datasets. However, these models are predominantly trained on large-scale semantic segmentation datasets, which introduce a bias toward object shape rather than texture cues in the image. This limitation is critical in domains such as medical imaging, material classification, and remote sensing, where texture changes define object boundaries. In this study, we investigate SAM's bias toward semantics over textures and introduce a new texture-aware foundation model, TextureSAM, which performs superior segmentation in texture-dominant scenarios. To achieve this, we employ a novel fine-tuning approach that incorporates texture augmentation techniques, incrementally modifying training images to emphasize texture features. By leveraging a novel texture-alternation of the ADE20K dataset, we guide TextureSAM to prioritize texture-defined regions, thereby mitigating the inherent shape bias present in the original SAM model. Our extensive experiments demonstrate that TextureSAM significantly outperforms SAM-2 on both natural (+0.2 mIoU) and synthetic (+0.18 mIoU) texture-based segmentation datasets. The code and texture-augmented dataset will be publicly available.

Matan Rusanovsky, Or Hirschorn, Shai Avidan

Conventional 2D pose estimation models are constrained by their design to specific object categories. This limits their applicability to predefined objects. To overcome these limitations, category-agnostic pose estimation (CAPE) emerged as a solution. CAPE aims to facilitate keypoint localization for diverse object categories using a unified model, which can generalize from minimal annotated support images. Recent CAPE works have produced object poses based on arbitrary keypoint definitions annotated on a user-provided support image. Our work departs from conventional CAPE methods, which require a support image, by adopting a text-based approach instead of the support image. Specifically, we use a pose-graph, where nodes represent keypoints that are described with text. This representation takes advantage of the abstraction of text descriptions and the structure imposed by the graph. Our approach effectively breaks symmetry, preserves structure, and improves occlusion handling. We validate our novel approach using the MP-100 benchmark, a comprehensive dataset spanning over 100 categories and 18,000 images. Under a 1-shot setting, our solution achieves a notable performance boost of 1.07\%, establishing a new state-of-the-art for CAPE. Additionally, we enrich the dataset by providing text description annotations, further enhancing its utility for future research.

Or Hirschorn, Amir Jevnisek, Shai Avidan

Vector image representation is a popular choice when editability and flexibility in resolution are desired. However, most images are only available in raster form, making raster-to-vector image conversion (vectorization) an important task. Classical methods for vectorization are either domain-specific or yield an abundance of shapes which limits editability and interpretability. Learning-based methods, that use differentiable rendering, have revolutionized vectorization, at the cost of poor generalization to out-of-training distribution domains, and optimization-based counterparts are either slow or produce non-editable and redundant shapes. In this work, we propose Optimize & Reduce (O&R), a top-down approach to vectorization that is both fast and domain-agnostic. O&R aims to attain a compact representation of input images by iteratively optimizing Bézier curve parameters and significantly reducing the number of shapes, using a devised importance measure. We contribute a benchmark of five datasets comprising images from a broad spectrum of image complexities - from emojis to natural-like images. Through extensive experiments on hundreds of images, we demonstrate that our method is domain agnostic and outperforms existing works in both reconstruction and perceptual quality for a fixed number of shapes. Moreover, we show that our algorithm is $\times 10$ faster than the state-of-the-art optimization-based method.

Daniel Kigli, Ariel Shamir, Shai Avidan

How many labeled pixels are needed to segment an image, without any prior knowledge? We conduct an experiment to answer this question. In our experiment, an Oracle is using Active Learning to train a network from scratch. The Oracle has access to the entire label map of the image, but the goal is to reveal as little pixel labels to the network as possible. We find that, on average, the Oracle needs to reveal (i.e., annotate) less than 0.1% of the pixels in order to train a network. The network can then label all pixels in the image at an accuracy of more than 98%. Based on this single-image-annotation experiment, we design an experiment to quickly annotate an entire data set. In the data set level experiment the Oracle trains a new network for each image from scratch. The network can then be used to create pseudo-labels, which are the network predicted labels of the unlabeled pixels, for the entire image. Only then, a data set level network is trained from scratch on all the pseudo-labeled images at once. We repeat both image level and data set level experiments on two, very different, real-world data sets, and find that it is possible to reach the performance of a fully annotated data set using a fraction of the annotation cost.

Alon Zolfi, Shai Avidan, Yuval Elovici et al.

Deep learning-based facial recognition (FR) models have demonstrated state-of-the-art performance in the past few years, even when wearing protective medical face masks became commonplace during the COVID-19 pandemic. Given the outstanding performance of these models, the machine learning research community has shown increasing interest in challenging their robustness. Initially, researchers presented adversarial attacks in the digital domain, and later the attacks were transferred to the physical domain. However, in many cases, attacks in the physical domain are conspicuous, and thus may raise suspicion in real-world environments (e.g., airports). In this paper, we propose Adversarial Mask, a physical universal adversarial perturbation (UAP) against state-of-the-art FR models that is applied on face masks in the form of a carefully crafted pattern. In our experiments, we examined the transferability of our adversarial mask to a wide range of FR model architectures and datasets. In addition, we validated our adversarial mask's effectiveness in real-world experiments (CCTV use case) by printing the adversarial pattern on a fabric face mask. In these experiments, the FR system was only able to identify 3.34% of the participants wearing the mask (compared to a minimum of 83.34% with other evaluated masks). A demo of our experiments can be found at: https://youtu.be/_TXkDO5z11w.

Itan Hezroni, Amnon Drory, Raja Giryes et al.

Recently, several deep learning approaches have been proposed for point cloud registration. These methods train a network to generate a representation that helps finding matching points in two 3D point clouds. Finding good matches allows them to calculate the transformation between the point clouds accurately. Two challenges of these techniques are dealing with occlusions and generalizing to objects of classes unseen during training. This work proposes DeepBBS, a novel method for learning a representation that takes into account the best buddy distance between points during training. Best Buddies (i.e., mutual nearest neighbors) are pairs of points nearest to each other. The Best Buddies criterion is a strong indication for correct matches that, in turn, leads to accurate registration. Our experiments show improved performance compared to previous methods. In particular, our learned representation leads to an accurate registration for partial shapes and in unseen categories.

Itzik Mizrahi, Shai Avidan

Deep Neural Networks require large amounts of labeled data for their training. Collecting this data at scale inevitably causes label noise.Hence,the need to develop learning algorithms that are robust to label noise. In recent years, k Nearest Neighbors (kNN) emerged as a viable solution to this problem. Despite its success, kNN is not without its problems. Mainly, it requires a huge memory footprint to store all the training samples and it needs an advanced data structure to allow for fast retrieval of the relevant examples, given a query sample. We propose a neural network, termed kNet, that learns to perform kNN. Once trained, we no longer need to store the training data, and processing a query sample is a simple matter of inference. To use kNet, we first train a preliminary network on the data set, and then train kNet on the penultimate layer of the preliminary network.We find that kNet gives a smooth approximation of kNN,and cannot handle the sharp label changes between samples that kNN can exhibit. This indicates that currently kNet is best suited to approximate kNN with a fairly large k. Experiments on two data sets show that this is the regime in which kNN works best,and can therefore be replaced by kNet.In practice, kNet consistently improve the results of all preliminary networks, in all label noise regimes, by up to 3%.

Inbar Helbitz, Shai Avidan

Privacy-Preserving Machine Learning algorithms must balance classification accuracy with data privacy. This can be done using a combination of cryptographic and machine learning tools such as Convolutional Neural Networks (CNN). CNNs typically consist of two types of operations: a convolutional or linear layer, followed by a non-linear function such as ReLU. Each of these types can be implemented efficiently using a different cryptographic tool. But these tools require different representations and switching between them is time-consuming and expensive. Recent research suggests that ReLU is responsible for most of the communication bandwidth. ReLU is usually applied at each pixel (or activation) location, which is quite expensive. We propose to share ReLU operations. Specifically, the ReLU decision of one activation can be used by others, and we explore different ways to group activations and different ways to determine the ReLU for such a group of activations. Experiments on several datasets reveal that we can cut the number of ReLU operations by up to three orders of magnitude and, as a result, cut the communication bandwidth by more than 50%.

Kfir Goldberg, Stav Shapiro, Elad Richardson et al.

The search for efficient neural network architectures has gained much focus in recent years, where modern architectures focus not only on accuracy but also on inference time and model size. Here, we present FUN, a family of novel Frequency-domain Utilization Networks. These networks utilize the inherent efficiency of the frequency-domain by working directly in that domain, represented with the Discrete Cosine Transform. Using modern techniques and building blocks such as compound-scaling and inverted-residual layers we generate a set of such networks allowing one to balance between size, latency and accuracy while outperforming competing RGB-based models. Extensive evaluations verifies that our networks present strong alternatives to previous approaches. Moreover, we show that working in frequency domain allows for dynamic compression of the input at inference time without any explicit change to the architecture.

Amnon Drory, Tal Shomer, Shai Avidan et al.

We propose new, and robust, loss functions for the point cloud registration problem. Our loss functions are inspired by the Best Buddies Similarity (BBS) measure that counts the number of mutual nearest neighbors between two point sets. This measure has been shown to be robust to outliers and missing data in the case of template matching for images. We present several algorithms, collectively named Best Buddy Registration (BBR), where each algorithm consists of optimizing one of these loss functions with Adam gradient descent. The loss functions differ in several ways, including the distance function used (point-to-point vs. point-to-plane), and how the BBS measure is combined with the actual distances between pairs of points. Experiments on various data sets, both synthetic and real, demonstrate the effectiveness of the BBR algorithms, showing that they are quite robust to noise, outliers, and distractors, and cope well with extremely sparse point clouds. One variant, BBR-F, achieves state-of-the-art accuracy in the registration of automotive lidar scans taken up to several seconds apart, from the KITTI and Apollo-Southbay datasets.

Anna Darzi, Itai Lang, Ashutosh Taklikar et al.

As image generation techniques mature, there is a growing interest in explainable representations that are easy to understand and intuitive to manipulate. In this work, we turn to co-occurrence statistics, which have long been used for texture analysis, to learn a controllable texture synthesis model. We propose a fully convolutional generative adversarial network, conditioned locally on co-occurrence statistics, to generate arbitrarily large images while having local, interpretable control over the texture appearance. To encourage fidelity to the input condition, we introduce a novel differentiable co-occurrence loss that is integrated seamlessly into our framework in an end-to-end fashion. We demonstrate that our solution offers a stable, intuitive and interpretable latent representation for texture synthesis, which can be used to generate a smooth texture morph between different textures. We further show an interactive texture tool that allows a user to adjust local characteristics of the synthesized texture image using the co-occurrence values directly.

Inbal Lav, Shai Avidan, Yoram Singer et al.

We introduce a binary embedding framework, called Proximity Preserving Code (PPC), which learns similarity and dissimilarity between data points to create a compact and affinity-preserving binary code. This code can be used to apply fast and memory-efficient approximation to nearest-neighbor searches. Our framework is flexible, enabling different proximity definitions between data points. In contrast to previous methods that extract binary codes based on unsigned graph partitioning, our system models the attractive and repulsive forces in the data by incorporating positive and negative graph weights. The proposed framework is shown to boil down to finding the minimal cut of a signed graph, a problem known to be NP-hard. We offer an efficient approximation and achieve superior results by constructing the code bit after bit. We show that the proposed approximation is superior to the commonly used spectral methods with respect to both accuracy and complexity. Thus, it is useful for many other problems that can be translated into signed graph cut.

Amir Markovitz, Gilad Sharir, Itamar Friedman et al.

We propose a new method for anomaly detection of human actions. Our method works directly on human pose graphs that can be computed from an input video sequence. This makes the analysis independent of nuisance parameters such as viewpoint or illumination. We map these graphs to a latent space and cluster them. Each action is then represented by its soft-assignment to each of the clusters. This gives a kind of "bag of words" representation to the data, where every action is represented by its similarity to a group of base action-words. Then, we use a Dirichlet process based mixture, that is useful for handling proportional data such as our soft-assignment vectors, to determine if an action is normal or not. We evaluate our method on two types of data sets. The first is a fine-grained anomaly detection data set (e.g. ShanghaiTech) where we wish to detect unusual variations of some action. The second is a coarse-grained anomaly detection data set (e.g., a Kinetics-based data set) where few actions are considered normal, and every other action should be considered abnormal. Extensive experiments on the benchmarks show that our method performs considerably better than other state of the art methods.

Matan Goldman, Tal Hassner, Shai Avidan

The field of self-supervised monocular depth estimation has seen huge advancements in recent years. Most methods assume stereo data is available during training but usually under-utilize it and only treat it as a reference signal. We propose a novel self-supervised approach which uses both left and right images equally during training, but can still be used with a single input image at test time, for monocular depth estimation. Our Siamese network architecture consists of two, twin networks, each learns to predict a disparity map from a single image. At test time, however, only one of these networks is used in order to infer depth. We show state-of-the-art results on the standard KITTI Eigen split benchmark as well as being the highest scoring self-supervised method on the new KITTI single view benchmark. To demonstrate the ability of our method to generalize to new data sets, we further provide results on the Make3D benchmark, which was not used during training.

Nir Zarrabi, Shai Avidan, Yael Moses

We consider the problem of segmenting dynamic regions in CrowdCam images, where a dynamic region is the projection of a moving 3D object on the image plane. Quite often, these regions are the most interesting parts of an image. CrowdCam images is a set of images of the same dynamic event, captured by a group of non-collaborating users. Almost every event of interest today is captured this way. This new type of images raises the need to develop new algorithms tailored specifically for it. We propose a comprehensive solution to the problem. Our solution combines cues that are based on geometry, appearance and proximity. First, geometric reasoning is used to produce rough score maps that determine, for every pixel, how likely it is to be the projection of a static or dynamic scene point. These maps are noisy because CrowdCam images are usually few and far apart both in space and in time. Then, we use similarity in appearance space and proximity in the image plane to encourage neighboring pixels to be labeled similarly as either static or dynamic. We collected a new, and challenging, data set to evaluate our algorithm. Results show that the success score of our algorithm is nearly double that of the current state of the art approach.

Dana Berman, Deborah Levy, Shai Avidan et al.

Underwater images suffer from color distortion and low contrast, because light is attenuated while it propagates through water. Attenuation under water varies with wavelength, unlike terrestrial images where attenuation is assumed to be spectrally uniform. The attenuation depends both on the water body and the 3D structure of the scene, making color restoration difficult. Unlike existing single underwater image enhancement techniques, our method takes into account multiple spectral profiles of different water types. By estimating just two additional global parameters: the attenuation ratios of the blue-red and blue-green color channels, the problem is reduced to single image dehazing, where all color channels have the same attenuation coefficients. Since the water type is unknown, we evaluate different parameters out of an existing library of water types. Each type leads to a different restored image and the best result is automatically chosen based on color distribution. We collected a dataset of images taken in different locations with varying water properties, showing color charts in the scenes. Moreover, to obtain ground truth, the 3D structure of the scene was calculated based on stereo imaging. This dataset enables a quantitative evaluation of restoration algorithms on natural images and shows the advantage of our method.

Amnon Drory, Oria Ratzon, Shai Avidan et al.

We investigate the classification performance of K-nearest neighbors (K-NN) and deep neural networks (DNNs) in the presence of label noise. We first show empirically that a DNN's prediction for a given test example depends on the labels of the training examples in its local neighborhood. This motivates us to derive a realizable analytic expression that approximates the multi-class K-NN classification error in the presence of label noise, which is of independent importance. We then suggest that the expression for K-NN may serve as a first-order approximation for the DNN error. Finally, we demonstrate empirically the proximity of the developed expression to the observed performance of K-NN and DNN classifiers. Our result may explain the already observed surprising resistance of DNN to some types of label noise. It also characterizes an important factor of it showing that the more concentrated the noise the greater is the degradation in performance.

Roy J Jevnisek, Shai Avidan

Co-occurrence Filter (CoF) is a boundary preserving filter. It is based on the Bilateral Filter (BF) but instead of using a Gaussian on the range values to preserve edges it relies on a co-occurrence matrix. Pixel values that co-occur frequently in the image (i.e., inside textured regions) will have a high weight in the co-occurrence matrix. This, in turn, means that such pixel pairs will be averaged and hence smoothed, regardless of their intensity differences. On the other hand, pixel values that rarely co-occur (i.e., across texture boundaries) will have a low weight in the co-occurrence matrix. As a result, they will not be averaged and the boundary between them will be preserved. The CoF therefore extends the BF to deal with boundaries, not just edges. It learns co-occurrences directly from the image. We can achieve various filtering results by directing it to learn the co-occurrence matrix from a part of the image, or a different image. We give the definition of the filter, discuss how to use it with color images and show several use cases.

Adi Dafni, Yael Moses, Shai Avidan

We address the novel problem of detecting dynamic regions in CrowdCam images, a set of still images captured by a group of people. These regions capture the most interesting parts of the scene, and detecting them plays an important role in the analysis of visual data. Our method is based on the observation that matching static points must satisfy the epipolar geometry constraints, but computing exact matches is challenging. Instead, we compute the probability that a pixel has a match, not necessarily the correct one, along the corresponding epipolar line. The complement of this probability is not necessarily the probability of a dynamic point because of occlusions, noise, and matching errors. Therefore, information from all pairs of images is aggregated to obtain a high quality dynamic probability map, per image. Experiments on challenging datasets demonstrate the effectiveness of the algorithm on a broad range of settings; no prior knowledge about the scene, the camera characteristics or the camera locations is required.

Shaul Oron, Denis Suhanov, Shai Avidan

Best-Buddies Tracking (BBT) applies the Best-Buddies Similarity measure (BBS) to the problem of model-free online tracking. BBS was introduced as a similarity measure between two point sets and was shown to be very effective for template matching. Originally, BBS was designed to work with point sets of equal size, and we propose a modification that lets it handle point sets of different size. The modified BBS is better suited to handle scale changes in the template size, as well as support a variable number of template images. We embed the modified BBS in a particle filter framework and obtain good results on a number of standard benchmarks.

Shaul Oron, Tali Dekel, Tianfan Xue et al.

We propose a novel method for template matching in unconstrained environments. Its essence is the Best-Buddies Similarity (BBS), a useful, robust, and parameter-free similarity measure between two sets of points. BBS is based on counting the number of Best-Buddies Pairs (BBPs)--pairs of points in source and target sets, where each point is the nearest neighbor of the other. BBS has several key features that make it robust against complex geometric deformations and high levels of outliers, such as those arising from background clutter and occlusions. We study these properties, provide a statistical analysis that justifies them, and demonstrate the consistent success of BBS on a challenging real-world dataset while using different types of features.

Tal Remez, Shai Avidan

Spatially Coherent Random Forest (SCRF) extends Random Forest to create spatially coherent labeling. Each split function in SCRF is evaluated based on a traditional information gain measure that is regularized by a spatial coherency term. This way, SCRF is encouraged to choose split functions that cluster pixels both in appearance space and in image space. In particular, we use SCRF to detect contours in images, where contours are taken to be the boundaries between different regions. Each tree in the forest produces a segmentation of the image plane and the boundaries of the segmentations of all trees are aggregated to produce a final hierarchical contour map. We show that this modification improves the performance of regular Random Forest by about 10% on the standard Berkeley Segmentation Datasets. We believe that SCRF can be used in other settings as well.

Dana Menaker, Shai Avidan

We propose an algorithm for recovering depth using less than two images. Instead of having both cameras send their entire image to the host computer, the left camera sends its image to the host while the right camera sends only a fraction $ε$ of its image. The key aspect is that the cameras send the information without communicating at all. Hence, the required communication bandwidth is significantly reduced. While standard image compression techniques can reduce the communication bandwidth, this requires additional computational resources on the part of the encoder (camera). We aim at designing a light weight encoder that only touches a fraction of the pixels. The burden of decoding is placed on the decoder (host). We show that it is enough for the encoder to transmit a sparse set of pixels. Using only $1+ε$ images, with $ε$ as little as 2% of the image, the decoder can compute a depth map. The depth map's accuracy is comparable to traditional stereo matching algorithms that require both images as input. Using the depth map and the left image, the right image can be synthesized. No computations are required at the encoder, and the decoder's runtime is linear in the images' size.