Computer Vision and Pattern Recognition (cs.CV)

Abstract: Domain shift is a commonly encountered issue in medical imaging solutions, primarily caused by variations in imaging devices and data sources. To mitigate this problem, unsupervised domain adaptation techniques have been employed. However, concerns regarding patient privacy and potential degradation of image quality have led to an increased focus on source-free domain adaptation. In this study, we address the issue of false labels in self-training based source-free domain adaptive medical image segmentation methods. To correct erroneous pseudo-labels, we propose a novel approach called the local-global pseudo-label correction (LGDA) method for source-free domain adaptive medical image segmentation. Our method consists of two components: An offline local context-based pseudo-label correction method that utilizes local context similarity in image space. And an online global pseudo-label correction method based on class prototypes, which corrects erroneously predicted pseudo-labels by considering the relative distance between pixel-wise feature vectors and prototype vectors. We evaluate the performance of our method on three benchmark fundus image datasets for optic disc and cup segmentation. Our method achieves superior performance compared to the state-of-the-art approaches, even without using of any source data.

Abstract: Fine-tuning pre-trained models has emerged as a powerful technique in numerous domains, owing to its ability to leverage enormous pre-existing knowledge and achieve remarkable performance on downstream tasks. However, updating the parameters of entire networks is computationally intensive. Although state-of-the-art parameter-efficient transfer learning (PETL) methods significantly reduce the trainable parameters and storage demand, almost all of them still need to back-propagate the gradients through large pre-trained networks. This memory-extensive characteristic extremely limits the applicability of PETL methods in real-world scenarios. To this end, we propose a new memory-efficient PETL strategy, dubbed Universal Parallel Tuning (UniPT). Specifically, we facilitate the transfer process via a lightweight learnable parallel network, which consists of two modules: 1) A parallel interaction module that decouples the inherently sequential connections and processes the intermediate activations detachedly of the pre-trained network. 2) A confidence aggregation module that learns optimal strategies adaptively for integrating cross-layer features. We evaluate UniPT with different backbones (e.g., VSE$\infty$, CLIP4Clip, Clip-ViL, and MDETR) on five challenging vision-and-language tasks (i.e., image-text retrieval, video-text retrieval, visual question answering, compositional question answering, and visual grounding). Extensive ablations on ten datasets have validated that our UniPT can not only dramatically reduce memory consumption and outperform the best memory-efficient competitor, but also achieve higher performance than existing PETL methods in a low-memory scenario on different architectures. Our code is publicly available at: https://github.com/Paranioar/UniPT.

Abstract: In recent years, the assessment of fundamental movement skills integrated with physical education has focused on both teaching practice and the feasibility of assessment. The object of assessment has shifted from multiple ages to subdivided ages, while the content of assessment has changed from complex and time-consuming to concise and efficient. Therefore, we apply deep learning to physical fitness evaluation, we propose a system based on the Canadian Agility and Movement Skill Assessment (CAMSA) Physical Fitness Evaluation System (CPFES), which evaluates children's physical fitness based on CAMSA, and gives recommendations based on the scores obtained by CPFES to help children grow. We have designed a landmark detection module and a pose estimation module, and we have also designed a pose evaluation module for the CAMSA criteria that can effectively evaluate the actions of the child being tested. Our experimental results demonstrate the high accuracy of the proposed system.

Abstract: LiDAR odometry estimation and 3D semantic segmentation are crucial for autonomous driving, which has achieved remarkable advances recently. However, these tasks are challenging due to the imbalance of points in different semantic categories for 3D semantic segmentation and the influence of dynamic objects for LiDAR odometry estimation, which increases the importance of using representative/salient landmarks as reference points for robust feature learning. To address these challenges, we propose a saliency-guided approach that leverages attention information to improve the performance of LiDAR odometry estimation and semantic segmentation models. Unlike in the image domain, only a few studies have addressed point cloud saliency information due to the lack of annotated training data. To alleviate this, we first present a universal framework to transfer saliency distribution knowledge from color images to point clouds, and use this to construct a pseudo-saliency dataset (i.e. FordSaliency) for point clouds. Then, we adopt point cloud-based backbones to learn saliency distribution from pseudo-saliency labels, which is followed by our proposed SalLiDAR module. SalLiDAR is a saliency-guided 3D semantic segmentation model that integrates saliency information to improve segmentation performance. Finally, we introduce SalLONet, a self-supervised saliency-guided LiDAR odometry network that uses the semantic and saliency predictions of SalLiDAR to achieve better odometry estimation. Our extensive experiments on benchmark datasets demonstrate that the proposed SalLiDAR and SalLONet models achieve state-of-the-art performance against existing methods, highlighting the effectiveness of image-to-LiDAR saliency knowledge transfer. Source code will be available at https://github.com/nevrez/SalLONet.

Abstract: Real-world vision tasks frequently suffer from the appearance of adverse weather conditions including rain, fog, snow, and raindrops in captured images. Recently, several generic methods for restoring weather-degraded images have been proposed, aiming to remove multiple types of adverse weather effects present in the images. However, these methods have considered weather as discrete and mutually exclusive variables, leading to failure in generalizing to unforeseen weather conditions beyond the scope of the training data, such as the co-occurrence of rain, fog, and raindrops. To this end, weather-degraded image restoration models should have flexible adaptability to the current unknown weather condition to ensure reliable and optimal performance. The adaptation method should also be able to cope with data scarcity for real-world adaptation. This paper proposes MetaWeather, a few-shot weather-degraded image restoration method for arbitrary weather conditions. For this, we devise the core piece of MetaWeather, coined Degradation Pattern Matching Module (DPMM), which leverages representations from a few-shot support set by matching features between input and sample images under new weather conditions. In addition, we build meta-knowledge with episodic meta-learning on top of our MetaWeather architecture to provide flexible adaptability. In the meta-testing phase, we adopt a parameter-efficient fine-tuning method to preserve the prebuilt knowledge and avoid the overfitting problem. Experiments on the BID Task II.A dataset show our method achieves the best performance on PSNR and SSIM compared to state-of-the-art image restoration methods. Code is available at (TBA).

Abstract: Learning-based surface reconstruction based on unsigned distance functions (UDF) has many advantages such as handling open surfaces. We propose SuperUDF, a self-supervised UDF learning which exploits a learned geometry prior for efficient training and a novel regularization for robustness to sparse sampling. The core idea of SuperUDF draws inspiration from the classical surface approximation operator of locally optimal projection (LOP). The key insight is that if the UDF is estimated correctly, the 3D points should be locally projected onto the underlying surface following the gradient of the UDF. Based on that, a number of inductive biases on UDF geometry and a pre-learned geometry prior are devised to learn UDF estimation efficiently. A novel regularization loss is proposed to make SuperUDF robust to sparse sampling. Furthermore, we also contribute a learning-based mesh extraction from the estimated UDFs. Extensive evaluations demonstrate that SuperUDF outperforms the state of the arts on several public datasets in terms of both quality and efficiency. Code will be released after accteptance.

Abstract: Multi-Label Image Recognition (MLIR) is a challenging task that aims to predict multiple object labels in a single image while modeling the complex relationships between labels and image regions. Although convolutional neural networks and vision transformers have succeeded in processing images as regular grids of pixels or patches, these representations are sub-optimal for capturing irregular and discontinuous regions of interest. In this work, we present the first fully graph convolutional model, Group K-nearest neighbor based Graph convolutional Network (GKGNet), which models the connections between semantic label embeddings and image patches in a flexible and unified graph structure. To address the scale variance of different objects and to capture information from multiple perspectives, we propose the Group KGCN module for dynamic graph construction and message passing. Our experiments demonstrate that GKGNet achieves state-of-the-art performance with significantly lower computational costs on the challenging multi-label datasets, \ie MS-COCO and VOC2007 datasets. We will release the code and models to facilitate future research in this area.

Abstract: Light-weight time-of-flight (ToF) depth sensors are compact and cost-efficient, and thus widely used on mobile devices for tasks such as autofocus and obstacle detection. However, due to the sparse and noisy depth measurements, these sensors have rarely been considered for dense geometry reconstruction. In this work, we present the first dense SLAM system with a monocular camera and a light-weight ToF sensor. Specifically, we propose a multi-modal implicit scene representation that supports rendering both the signals from the RGB camera and light-weight ToF sensor which drives the optimization by comparing with the raw sensor inputs. Moreover, in order to guarantee successful pose tracking and reconstruction, we exploit a predicted depth as an intermediate supervision and develop a coarse-to-fine optimization strategy for efficient learning of the implicit representation. At last, the temporal information is explicitly exploited to deal with the noisy signals from light-weight ToF sensors to improve the accuracy and robustness of the system. Experiments demonstrate that our system well exploits the signals of light-weight ToF sensors and achieves competitive results both on camera tracking and dense scene reconstruction. Project page: \url{https://zju3dv.github.io/tof_slam/}.

Abstract: Food computing has emerged as a prominent multidisciplinary field of research in recent years. An ambitious goal of food computing is to develop end-to-end intelligent systems capable of autonomously producing recipe information for a food image. Current image-to-recipe methods are retrieval-based and their success depends heavily on the dataset size and diversity, as well as the quality of learned embeddings. Meanwhile, the emergence of powerful attention-based vision and language models presents a promising avenue for accurate and generalizable recipe generation, which has yet to be extensively explored. This paper proposes FIRE, a novel multimodal methodology tailored to recipe generation in the food computing domain, which generates the food title, ingredients, and cooking instructions based on input food images. FIRE leverages the BLIP model to generate titles, utilizes a Vision Transformer with a decoder for ingredient extraction, and employs the T5 model to generate recipes incorporating titles and ingredients as inputs. We showcase two practical applications that can benefit from integrating FIRE with large language model prompting: recipe customization to fit recipes to user preferences and recipe-to-code transformation to enable automated cooking processes. Our experimental findings validate the efficacy of our proposed approach, underscoring its potential for future advancements and widespread adoption in food computing.

Abstract: Video instance segmentation is a challenging task that serves as the cornerstone of numerous downstream applications, including video editing and autonomous driving. In this report, we present further improvements to the SOTA VIS method, DVIS. First, we introduce a denoising training strategy for the trainable tracker, allowing it to achieve more stable and accurate object tracking in complex and long videos. Additionally, we explore the role of visual foundation models in video instance segmentation. By utilizing a frozen VIT-L model pre-trained by DINO v2, DVIS demonstrates remarkable performance improvements. With these enhancements, our method achieves 57.9 AP and 56.0 AP in the development and test phases, respectively, and ultimately ranked 1st in the VIS track of the 5th LSVOS Challenge. The code will be available at https://github.com/zhang-tao-whu/DVIS.

Abstract: In this research paper, we present an innovative system designed for the purpose of segmenting the layout of Bangla documents. Our methodology involves utilizing a sophisticated collection of YOLOv8 models, meticulously adapted for the DL Sprint 2.0 - BUET CSE Fest 2023 Competition that centers around Bangla document layout segmentation. Our primary focus lies in elevating various elements of the task, including techniques like image augmentation, model architecture, and the use of model ensembles. We intentionally lower the quality of a subset of document images to enhance the resilience of model training, consequently leading to an improvement in our cross-validation score. Employing Bayesian optimization, we determine the optimal confidence and IoU thresholds for our model ensemble. Through our approach, we successfully showcase the effectiveness of amalgamating anchor-free models to achieve robust layout segmentation in Bangla documents.

Abstract: In computer vision, depth estimation is crucial for domains like robotics, autonomous vehicles, augmented reality, and virtual reality. Integrating semantics with depth enhances scene understanding through reciprocal information sharing. However, the scarcity of semantic information in datasets poses challenges. Existing convolutional approaches with limited local receptive fields hinder the full utilization of the symbiotic potential between depth and semantics. This paper introduces a dataset-invariant semi-supervised strategy to address the scarcity of semantic information. It proposes the Depth Semantics Symbiosis module, leveraging the Symbiotic Transformer for achieving comprehensive mutual awareness by information exchange within both local and global contexts. Additionally, a novel augmentation, NearFarMix is introduced to combat overfitting and compensate both depth-semantic tasks by strategically merging regions from two images, generating diverse and structurally consistent samples with enhanced control. Extensive experiments on NYU-Depth-V2 and KITTI datasets demonstrate the superiority of our proposed techniques in indoor and outdoor environments.

Abstract: Denoising diffusion models have emerged as the go-to framework for solving inverse problems in imaging. A critical concern regarding these models is their performance on out-of-distribution (OOD) tasks, which remains an under-explored challenge. Realistic reconstructions inconsistent with the measured data can be generated, hallucinating image features that are uniquely present in the training dataset. To simultaneously enforce data-consistency and leverage data-driven priors, we introduce a novel sampling framework called Steerable Conditional Diffusion. This framework adapts the denoising network specifically to the available measured data. Utilising our proposed method, we achieve substantial enhancements in OOD performance across diverse imaging modalities, advancing the robust deployment of denoising diffusion models in real-world applications.

Abstract: Sensor fusion has become a popular topic in robotics. However, conventional fusion methods encounter many difficulties, such as data representation differences, sensor variations, and extrinsic calibration. For example, the calibration methods used for LiDAR-camera fusion often require manual operation and auxiliary calibration targets. Implicit neural representations (INRs) have been developed for 3D scenes, and the volume density distribution involved in an INR unifies the scene information obtained by different types of sensors. Therefore, we propose implicit neural fusion (INF) for LiDAR and camera. INF first trains a neural density field of the target scene using LiDAR frames. Then, a separate neural color field is trained using camera images and the trained neural density field. Along with the training process, INF both estimates LiDAR poses and optimizes extrinsic parameters. Our experiments demonstrate the high accuracy and stable performance of the proposed method.

Abstract: During the past decade, deep neural networks have led to fast-paced progress and significant achievements in computer vision problems, for both academia and industry. Yet despite their success, state-of-the-art image classification approaches fail to generalize well in previously unseen visual contexts, as required by many real-world applications. In this paper, we focus on this domain generalization (DG) problem and argue that the generalization ability of deep convolutional neural networks can be improved by taking advantage of multi-layer and multi-scaled representations of the network. We introduce a framework that aims at improving domain generalization of image classifiers by combining both low-level and high-level features at multiple scales, enabling the network to implicitly disentangle representations in its latent space and learn domain-invariant attributes of the depicted objects. Additionally, to further facilitate robust representation learning, we propose a novel objective function, inspired by contrastive learning, which aims at constraining the extracted representations to remain invariant under distribution shifts. We demonstrate the effectiveness of our method by evaluating on the domain generalization datasets of PACS, VLCS, Office-Home and NICO. Through extensive experimentation, we show that our model is able to surpass the performance of previous DG methods and consistently produce competitive and state-of-the-art results in all datasets.

Abstract: Different from traditional video cameras, event cameras capture asynchronous events stream in which each event encodes pixel location, trigger time, and the polarity of the brightness changes. In this paper, we introduce a novel graph-based framework for event cameras, namely SlideGCN. Unlike some recent graph-based methods that use groups of events as input, our approach can efficiently process data event-by-event, unlock the low latency nature of events data while still maintaining the graph's structure internally. For fast graph construction, we develop a radius search algorithm, which better exploits the partial regular structure of event cloud against k-d tree based generic methods. Experiments show that our method reduces the computational complexity up to 100 times with respect to current graph-based methods while keeping state-of-the-art performance on object recognition. Moreover, we verify the superiority of event-wise processing with our method. When the state becomes stable, we can give a prediction with high confidence, thus making an early recognition. Project page: \url{https://zju3dv.github.io/slide_gcn/}.

Abstract: Score-based generative models (SGMs) have gained prominence in sparse-view CT reconstruction for their precise sampling of complex distributions. In SGM-based reconstruction, data consistency in the score-based diffusion model ensures close adherence of generated samples to observed data distribution, crucial for improving image quality. Shortcomings in data consistency characterization manifest in three aspects. Firstly, data from the optimization process can lead to artifacts in reconstructed images. Secondly, it often neglects that the generation model and original data constraints are independently completed, fragmenting unity. Thirdly, it predominantly focuses on constraining intermediate results in the inverse sampling process, rather than ideal real images. Thus, we propose an iterative optimization data scoring model. This paper introduces the data-iterative optimization score-based model (DOSM), integrating innovative data consistency into the Stochastic Differential Equation, a valuable constraint for ultra-sparse-view CT reconstruction. The novelty of this data consistency element lies in its sole reliance on original measurement data to confine generation outcomes, effectively balancing measurement data and generative model constraints. Additionally, we pioneer an inference strategy that traces back from current iteration results to ideal truth, enhancing reconstruction stability. We leverage conventional iteration techniques to optimize DOSM updates. Quantitative and qualitative results from 23 views of numerical and clinical cardiac datasets demonstrate DOSM's superiority over other methods. Remarkably, even with 10 views, our method achieves excellent performance.

Abstract: The objective of stylized speech-driven facial animation is to create animations that encapsulate specific emotional expressions. Existing methods often depend on pre-established emotional labels or facial expression templates, which may limit the necessary flexibility for accurately conveying user intent. In this research, we introduce a technique that enables the control of arbitrary styles by leveraging natural language as emotion prompts. This technique presents benefits in terms of both flexibility and user-friendliness. To realize this objective, we initially construct a Text-Expression Alignment Dataset (TEAD), wherein each facial expression is paired with several prompt-like descriptions.We propose an innovative automatic annotation method, supported by Large Language Models (LLMs), to expedite the dataset construction, thereby eliminating the substantial expense of manual annotation. Following this, we utilize TEAD to train a CLIP-based model, termed ExpCLIP, which encodes text and facial expressions into semantically aligned style embeddings. The embeddings are subsequently integrated into the facial animation generator to yield expressive and controllable facial animations. Given the limited diversity of facial emotions in existing speech-driven facial animation training data, we further introduce an effective Expression Prompt Augmentation (EPA) mechanism to enable the animation generator to support unprecedented richness in style control. Comprehensive experiments illustrate that our method accomplishes expressive facial animation generation and offers enhanced flexibility in effectively conveying the desired style.

Abstract: Over the last ten years, Patient-Derived Organoids (PDOs) emerged as the most reliable technology to generate ex-vivo tumor avatars. PDOs retain the main characteristics of their original tumor, making them a system of choice for pre-clinical and clinical studies. In particular, PDOs are attracting interest in the field of Functional Precision Medicine (FPM), which is based upon an ex-vivo drug test in which living tumor cells (such as PDOs) from a specific patient are exposed to a panel of anti-cancer drugs. Currently, the Adenosine Triphosphate (ATP) based cell viability assay is the gold standard test to assess the sensitivity of PDOs to drugs. The readout is measured at the end of the assay from a global PDO population and therefore does not capture single PDO responses and does not provide time resolution of drug effect. To this end, in this study, we explore for the first time the use of powerful large foundation models for the automatic processing of PDO data. In particular, we propose a novel imaging-based high-throughput screening method to assess real-time drug efficacy from a time-lapse microscopy video of PDOs. The recently proposed SAM algorithm for segmentation and DINOv2 model are adapted in a comprehensive pipeline for processing PDO microscopy frames. Moreover, an attention mechanism is proposed for fusing temporal and spatial features in a multiple instance learning setting to predict ATP. We report better results than other non-time-resolved methods, indicating that the temporality of data is an important factor for the prediction of ATP. Extensive ablations shed light on optimizing the experimental setting and automating the prediction both in real-time and for forecasting.

Abstract: Object detection is a task that performs position identification and label classification of objects in images or videos. The information obtained through this process plays an essential role in various tasks in the field of computer vision. In object detection, the data utilized for training and validation typically originate from public datasets that are well-balanced in terms of the number of objects ascribed to each class in an image. However, in real-world scenarios, handling datasets with much greater class imbalance, i.e., very different numbers of objects for each class , is much more common, and this imbalance may reduce the performance of object detection when predicting unseen test images. In our study, thus, we propose a method that evenly distributes the classes in an image for training and validation, solving the class imbalance problem in object detection. Our proposed method aims to maintain a uniform class distribution through multi-label stratification. We tested our proposed method not only on public datasets that typically exhibit balanced class distribution but also on custom datasets that may have imbalanced class distribution. We found that our proposed method was more effective on datasets containing severe imbalance and less data. Our findings indicate that the proposed method can be effectively used on datasets with substantially imbalanced class distribution.

Abstract: Realistic image super-resolution (Real-ISR) aims to reproduce perceptually realistic image details from a low-quality input. The commonly used adversarial training based Real-ISR methods often introduce unnatural visual artifacts and fail to generate realistic textures for natural scene images. The recently developed generative stable diffusion models provide a potential solution to Real-ISR with pre-learned strong image priors. However, the existing methods along this line either fail to keep faithful pixel-wise image structures or resort to extra skipped connections to reproduce details, which requires additional training in image space and limits their extension to other related tasks in latent space such as image stylization. In this work, we propose a pixel-aware stable diffusion (PASD) network to achieve robust Real-ISR as well as personalized stylization. In specific, a pixel-aware cross attention module is introduced to enable diffusion models perceiving image local structures in pixel-wise level, while a degradation removal module is used to extract degradation insensitive features to guide the diffusion process together with image high level information. By simply replacing the base diffusion model with a personalized one, our method can generate diverse stylized images without the need to collect pairwise training data. PASD can be easily integrated into existing diffusion models such as Stable Diffusion. Experiments on Real-ISR and personalized stylization demonstrate the effectiveness of our proposed approach. The source code and models can be found at \url{https://github.com/yangxy/PASD}.

Abstract: Deep needle insertion to a target often poses a huge challenge, requiring a combination of specialized skills, assistive technology, and extensive training. One of the frequently encountered medical scenarios demanding such expertise includes the needle insertion into a femoral vessel in the groin. After the access to the femoral vessel, various medical procedures, such as cardiac catheterization and extracorporeal membrane oxygenation (ECMO) can be performed. However, even with the aid of Ultrasound imaging, achieving successful insertion can necessitate multiple attempts due to the complexities of anatomy and tissue deformation. To address this challenge, this paper presents an innovative technology for needle tip real-time tracking, aiming for enhanced needle insertion guidance. Specifically, our approach revolves around the creation of scattering imaging using an optical fiber-equipped needle, and uses Convolutional Neural Network (CNN) based algorithms to enable real-time estimation of the needle tip's position and orientation during insertion procedures. The efficacy of the proposed technology was rigorously evaluated through three experiments. The first two experiments involved rubber and bacon phantoms to simulate groin anatomy. The positional errors averaging 2.3+1.5mm and 2.0+1.2mm, and the orientation errors averaging 0.2+0.11rad and 0.16+0.1rad. Furthermore, the system's capabilities were validated through experiments conducted on fresh porcine phantom mimicking more complex anatomical structures, yielding positional accuracy results of 3.2+3.1mm and orientational accuracy of 0.19+0.1rad. Given the average femoral arterial radius of 4 to 5mm, the proposed system is demonstrated with a great potential for precise needle guidance in femoral artery insertion procedures. In addition, the findings highlight the broader potential applications of the system in the medical field.

Abstract: Text-to-motion generation is a formidable task, aiming to produce human motions that align with the input text while also adhering to human capabilities and physical laws. While there have been advancements in diffusion models, their application in discrete spaces remains underexplored. Current methods often overlook the varying significance of different motions, treating them uniformly. It is essential to recognize that not all motions hold the same relevance to a particular textual description. Some motions, being more salient and informative, should be given precedence during generation. In response, we introduce a Priority-Centric Motion Discrete Diffusion Model (M2DM), which utilizes a Transformer-based VQ-VAE to derive a concise, discrete motion representation, incorporating a global self-attention mechanism and a regularization term to counteract code collapse. We also present a motion discrete diffusion model that employs an innovative noise schedule, determined by the significance of each motion token within the entire motion sequence. This approach retains the most salient motions during the reverse diffusion process, leading to more semantically rich and varied motions. Additionally, we formulate two strategies to gauge the importance of motion tokens, drawing from both textual and visual indicators. Comprehensive experiments on the HumanML3D and KIT-ML datasets confirm that our model surpasses existing techniques in fidelity and diversity, particularly for intricate textual descriptions.

Abstract: Human pose and shape estimation (HPS) has attracted increasing attention in recent years. While most existing studies focus on HPS from 2D images or videos with inherent depth ambiguity, there are surging need to investigate HPS from 3D point clouds as depth sensors have been frequently employed in commercial devices. However, real-world sensory 3D points are usually noisy and incomplete, and also human bodies could have different poses of high diversity. To tackle these challenges, we propose a principled framework, PointHPS, for accurate 3D HPS from point clouds captured in real-world settings, which iteratively refines point features through a cascaded architecture. Specifically, each stage of PointHPS performs a series of downsampling and upsampling operations to extract and collate both local and global cues, which are further enhanced by two novel modules: 1) Cross-stage Feature Fusion (CFF) for multi-scale feature propagation that allows information to flow effectively through the stages, and 2) Intermediate Feature Enhancement (IFE) for body-aware feature aggregation that improves feature quality after each stage. To facilitate a comprehensive study under various scenarios, we conduct our experiments on two large-scale benchmarks, comprising i) a dataset that features diverse subjects and actions captured by real commercial sensors in a laboratory environment, and ii) controlled synthetic data generated with realistic considerations such as clothed humans in crowded outdoor scenes. Extensive experiments demonstrate that PointHPS, with its powerful point feature extraction and processing scheme, outperforms State-of-the-Art methods by significant margins across the board. Homepage: https://caizhongang.github.io/projects/PointHPS/.

Abstract: Skeleton-based action segmentation requires recognizing composable actions in untrimmed videos. Current approaches decouple this problem by first extracting local visual features from skeleton sequences and then processing them by a temporal model to classify frame-wise actions. However, their performances remain limited as the visual features cannot sufficiently express composable actions. In this context, we propose Latent Action Composition (LAC), a novel self-supervised framework aiming at learning from synthesized composable motions for skeleton-based action segmentation. LAC is composed of a novel generation module towards synthesizing new sequences. Specifically, we design a linear latent space in the generator to represent primitive motion. New composed motions can be synthesized by simply performing arithmetic operations on latent representations of multiple input skeleton sequences. LAC leverages such synthesized sequences, which have large diversity and complexity, for learning visual representations of skeletons in both sequence and frame spaces via contrastive learning. The resulting visual encoder has a high expressive power and can be effectively transferred onto action segmentation tasks by end-to-end fine-tuning without the need for additional temporal models. We conduct a study focusing on transfer-learning and we show that representations learned from pre-trained LAC outperform the state-of-the-art by a large margin on TSU, Charades, PKU-MMD datasets.

Abstract: Visual bird's eye view (BEV) semantic segmentation helps autonomous vehicles understand the surrounding environment only from images, including static elements (e.g., roads) and dynamic elements (e.g., vehicles, pedestrians). However, the high cost of annotation procedures of full-supervised methods limits the capability of the visual BEV semantic segmentation, which usually needs HD maps, 3D object bounding boxes, and camera extrinsic matrixes. In this paper, we present a novel semi-supervised framework for visual BEV semantic segmentation to boost performance by exploiting unlabeled images during the training. A consistency loss that makes full use of unlabeled data is then proposed to constrain the model on not only semantic prediction but also the BEV feature. Furthermore, we propose a novel and effective data augmentation method named conjoint rotation which reasonably augments the dataset while maintaining the geometric relationship between the front-view images and the BEV semantic segmentation. Extensive experiments on the nuScenes and Argoverse datasets show that our semi-supervised framework can effectively improve prediction accuracy. To the best of our knowledge, this is the first work that explores improving visual BEV semantic segmentation performance using unlabeled data. The code will be publicly available.

Abstract: Recent face presentation attack detection (PAD) leverages domain adaptation (DA) and domain generalization (DG) techniques to address performance degradation on unknown domains. However, DA-based PAD methods require access to unlabeled target data, while most DG-based PAD solutions rely on a priori, i.e., known domain labels. Moreover, most DA-/DG-based methods are computationally intensive, demanding complex model architectures and/or multi-stage training processes. This paper proposes to model face PAD as a compound DG task from a causal perspective, linking it to model optimization. We excavate the causal factors hidden in the high-level representation via counterfactual intervention. Moreover, we introduce a class-guided MixStyle to enrich feature-level data distribution within classes instead of focusing on domain information. Both class-guided MixStyle and counterfactual intervention components introduce no extra trainable parameters and negligible computational resources. Extensive cross-dataset and analytic experiments demonstrate the effectiveness and efficiency of our method compared to state-of-the-art PADs. The implementation and the trained weights are publicly available.

Abstract: Change detection (CD) is a fundamental and important task for monitoring the land surface dynamics in the earth observation field. Existing deep learning-based CD methods typically extract bi-temporal image features using a weight-sharing Siamese encoder network and identify change regions using a decoder network. These CD methods, however, still perform far from satisfactorily as we observe that 1) deep encoder layers focus on irrelevant background regions and 2) the models' confidence in the change regions is inconsistent at different decoder stages. The first problem is because deep encoder layers cannot effectively learn from imbalanced change categories using the sole output supervision, while the second problem is attributed to the lack of explicit semantic consistency preservation. To address these issues, we design a novel similarity-aware attention flow network (SAAN). SAAN incorporates a similarity-guided attention flow module with deeply supervised similarity optimization to achieve effective change detection. Specifically, we counter the first issue by explicitly guiding deep encoder layers to discover semantic relations from bi-temporal input images using deeply supervised similarity optimization. The extracted features are optimized to be semantically similar in the unchanged regions and dissimilar in the changing regions. The second drawback can be alleviated by the proposed similarity-guided attention flow module, which incorporates similarity-guided attention modules and attention flow mechanisms to guide the model to focus on discriminative channels and regions. We evaluated the effectiveness and generalization ability of the proposed method by conducting experiments on a wide range of CD tasks. The experimental results demonstrate that our method achieves excellent performance on several CD tasks, with discriminative features and semantic consistency preserved.

Abstract: Existing Referring Image Segmentation (RIS) methods typically require expensive pixel-level or box-level annotations for supervision. In this paper, we observe that the referring texts used in RIS already provide sufficient information to localize the target object. Hence, we propose a novel weakly-supervised RIS framework to formulate the target localization problem as a classification process to differentiate between positive and negative text expressions. While the referring text expressions for an image are used as positive expressions, the referring text expressions from other images can be used as negative expressions for this image. Our framework has three main novelties. First, we propose a bilateral prompt method to facilitate the classification process, by harmonizing the domain discrepancy between visual and linguistic features. Second, we propose a calibration method to reduce noisy background information and improve the correctness of the response maps for target object localization. Third, we propose a positive response map selection strategy to generate high-quality pseudo-labels from the enhanced response maps, for training a segmentation network for RIS inference. For evaluation, we propose a new metric to measure localization accuracy. Experiments on four benchmarks show that our framework achieves promising performances to existing fully-supervised RIS methods while outperforming state-of-the-art weakly-supervised methods adapted from related areas. Code is available at https://github.com/fawnliu/TRIS.

Abstract: Manually reading and logging gauge data is time inefficient, and the effort increases according to the number of gauges available. We present a computer vision pipeline that automates the reading of analog gauges. We propose a two-stage CNN pipeline that identifies the key structural components of an analog gauge and outputs an angular reading. To facilitate the training of our approach, a synthetic dataset is generated thus obtaining a set of realistic analog gauges with their corresponding annotation. To validate our proposal, an additional real-world dataset was collected with 4.813 manually curated images. When compared against state-of-the-art methodologies, our method shows a significant improvement of 4.55 in the average error, which is a 52% relative improvement. The resources for this project will be made available at: https://github.com/fuankarion/automatic-gauge-reading.

Abstract: Unsupervised anomaly detection (UAD) is a widely adopted approach in industry due to rare anomaly occurrences and data imbalance. A desirable characteristic of an UAD model is contained generalization ability which excels in the reconstruction of seen normal patterns but struggles with unseen anomalies. Recent studies have pursued to contain the generalization capability of their UAD models in reconstruction from different perspectives, such as design of neural network (NN) structure and training strategy. In contrast, we note that containing of generalization ability in reconstruction can also be obtained simply from steep-shaped loss landscape. Motivated by this, we propose a loss landscape sharpening method by amplifying the reconstruction loss, dubbed Loss AMPlification (LAMP). LAMP deforms the loss landscape into a steep shape so the reconstruction error on unseen anomalies becomes greater. Accordingly, the anomaly detection performance is improved without any change of the NN architecture. Our findings suggest that LAMP can be easily applied to any reconstruction error metrics in UAD settings where the reconstruction model is trained with anomaly-free samples only.

Abstract: Despite significant advances in deep learning, models often struggle to generalize well to new, unseen domains, especially when training data is limited. To address this challenge, we propose a novel approach for distribution-aware latent augmentation that leverages the relationships across samples to guide the augmentation procedure. Our approach first degrades the samples stochastically in the latent space, mapping them to augmented labels, and then restores the samples from their corrupted versions during training. This process confuses the classifier in the degradation step and restores the overall class distribution of the original samples, promoting diverse intra-class/cross-domain variability. We extensively evaluate our approach on a diverse set of datasets and tasks, including domain generalization benchmarks and medical imaging datasets with strong domain shift, where we show our approach achieves significant improvements over existing methods for latent space augmentation. We further show that our method can be flexibly adapted to long-tail recognition tasks, demonstrating its versatility in building more generalizable models. Code is available at https://github.com/nerdslab/LatentDR.

Abstract: Rapid progress is being made in developing large, pretrained, task-agnostic foundational vision models such as CLIP, ALIGN, DINOv2, etc. In fact, we are approaching the point where these models do not have to be finetuned downstream, and can simply be used in zero-shot or with a lightweight probing head. Critically, given the complexity of working at this scale, there is a bottleneck where relatively few organizations in the world are executing the training then sharing the models on centralized platforms such as HuggingFace and torch.hub. The goal of this work is to identify several key adversarial vulnerabilities of these models in an effort to make future designs more robust. Intuitively, our attacks manipulate deep feature representations to fool an out-of-distribution (OOD) detector which will be required when using these open-world-aware models to solve closed-set downstream tasks. Our methods reliably make in-distribution (ID) images (w.r.t. a downstream task) be predicted as OOD and vice versa while existing in extremely low-knowledge-assumption threat models. We show our attacks to be potent in whitebox and blackbox settings, as well as when transferred across foundational model types (e.g., attack DINOv2 with CLIP)! This work is only just the beginning of a long journey towards adversarially robust foundational vision models.

Abstract: In this work we introduce S-TREK, a novel local feature extractor that combines a deep keypoint detector, which is both translation and rotation equivariant by design, with a lightweight deep descriptor extractor. We train the S-TREK keypoint detector within a framework inspired by reinforcement learning, where we leverage a sequential procedure to maximize a reward directly related to keypoint repeatability. Our descriptor network is trained following a "detect, then describe" approach, where the descriptor loss is evaluated only at those locations where keypoints have been selected by the already trained detector. Extensive experiments on multiple benchmarks confirm the effectiveness of our proposed method, with S-TREK often outperforming other state-of-the-art methods in terms of repeatability and quality of the recovered poses, especially when dealing with in-plane rotations.

Abstract: Segment Anything Model (SAM) has received remarkable attention as it offers a powerful and versatile solution for object segmentation in images. However, fine-tuning SAM for downstream segmentation tasks under different scenarios remains a challenge, as the varied characteristics of different scenarios naturally requires diverse model parameter spaces. Most existing fine-tuning methods attempt to bridge the gaps among different scenarios by introducing a set of new parameters to modify SAM's original parameter space. Unlike these works, in this paper, we propose fine-tuning SAM efficiently by parameter space reconstruction (SAM-PARSER), which introduce nearly zero trainable parameters during fine-tuning. In SAM-PARSER, we assume that SAM's original parameter space is relatively complete, so that its bases are able to reconstruct the parameter space of a new scenario. We obtain the bases by matrix decomposition, and fine-tuning the coefficients to reconstruct the parameter space tailored to the new scenario by an optimal linear combination of the bases. Experimental results show that SAM-PARSER exhibits superior segmentation performance across various scenarios, while reducing the number of trainable parameters by $\approx 290$ times compared with current parameter-efficient fine-tuning methods.

Abstract: Deep learning harnesses massive parallel floating-point processing to train and evaluate large neural networks. Trends indicate that deeper and larger neural networks with an increasing number of parameters achieve higher accuracy than smaller neural networks. This performance improvement, which often requires heavy compute for both training and evaluation, eventually needs to translate well to resource-constrained hardware for practical value. Structured pruning asserts that while large networks enable us to find solutions to complex computer vision problems, a smaller, computationally efficient sub-network can be derived from the large neural network that retains model accuracy but significantly improves computational efficiency. We generalize structured pruning with algorithms for network augmentation, pruning, sub-network collapse and removal. In addition, we demonstrate efficient and stable convergence up to 93% sparsity and 95% FLOPs reduction without loss of inference accuracy using with continuous relaxation matching or exceeding the state of the art for all structured pruning methods. The resulting CNN executes efficiently on GPU hardware without computationally expensive sparse matrix operations. We achieve this with routine automatable operations on classification and segmentation problems using CIFAR-10, ImageNet, and CityScapes datasets with the ResNet and U-NET network architectures.

Abstract: Synthetic aperture radar (SAR) imaging technology is commonly used to provide 24-hour all-weather earth observation. However, it still has some drawbacks in SAR target classification, especially in fine-grained classification of aircraft: aircrafts in SAR images have large intra-class diversity and inter-class similarity; the number of effective samples is insufficient and it's hard to annotate. To address these issues, this article proposes a novel multi-modal self-supervised network (MS-Net) for fine-grained classification of aircraft. Firstly, in order to entirely exploit the potential of multi-modal information, a two-sided path feature extraction network (TSFE-N) is constructed to enhance the image feature of the target and obtain the domain knowledge feature of text mode. Secondly, a contrastive self-supervised learning (CSSL) framework is employed to effectively learn useful label-independent feature from unbalanced data, a similarity per-ception loss (SPloss) is proposed to avoid network overfitting. Finally, TSFE-N is used as the encoder of CSSL to obtain the classification results. Through a large number of experiments, our MS-Net can effectively reduce the difficulty of classifying similar types of aircrafts. In the case of no label, the proposed algorithm achieves an accuracy of 88.46% for 17 types of air-craft classification task, which has pioneering significance in the field of fine-grained classification of aircraft in SAR images.

Abstract: In stark contrast to the case of images, finding a concise, learnable discrete representation of 3D surfaces remains a challenge. In particular, while polygon meshes are arguably the most common surface representation used in geometry processing, their irregular and combinatorial structure often make them unsuitable for learning-based applications. In this work, we present VoroMesh, a novel and differentiable Voronoi-based representation of watertight 3D shape surfaces. From a set of 3D points (called generators) and their associated occupancy, we define our boundary representation through the Voronoi diagram of the generators as the subset of Voronoi faces whose two associated (equidistant) generators are of opposite occupancy: the resulting polygon mesh forms a watertight approximation of the target shape's boundary. To learn the position of the generators, we propose a novel loss function, dubbed VoroLoss, that minimizes the distance from ground truth surface samples to the closest faces of the Voronoi diagram which does not require an explicit construction of the entire Voronoi diagram. A direct optimization of the Voroloss to obtain generators on the Thingi32 dataset demonstrates the geometric efficiency of our representation compared to axiomatic meshing algorithms and recent learning-based mesh representations. We further use VoroMesh in a learning-based mesh prediction task from input SDF grids on the ABC dataset, and show comparable performance to state-of-the-art methods while guaranteeing closed output surfaces free of self-intersections.

Abstract: Deep-learning models for 3D point cloud semantic segmentation exhibit limited generalization capabilities when trained and tested on data captured with different sensors or in varying environments due to domain shift. Domain adaptation methods can be employed to mitigate this domain shift, for instance, by simulating sensor noise, developing domain-agnostic generators, or training point cloud completion networks. Often, these methods are tailored for range view maps or necessitate multi-modal input. In contrast, domain adaptation in the image domain can be executed through sample mixing, which emphasizes input data manipulation rather than employing distinct adaptation modules. In this study, we introduce compositional semantic mixing for point cloud domain adaptation, representing the first unsupervised domain adaptation technique for point cloud segmentation based on semantic and geometric sample mixing. We present a two-branch symmetric network architecture capable of concurrently processing point clouds from a source domain (e.g. synthetic) and point clouds from a target domain (e.g. real-world). Each branch operates within one domain by integrating selected data fragments from the other domain and utilizing semantic information derived from source labels and target (pseudo) labels. Additionally, our method can leverage a limited number of human point-level annotations (semi-supervised) to further enhance performance. We assess our approach in both synthetic-to-real and real-to-real scenarios using LiDAR datasets and demonstrate that it significantly outperforms state-of-the-art methods in both unsupervised and semi-supervised settings.

Abstract: BACKGROUND & AIMS: Histological remission (HR) is advocated and considered as a new therapeutic target in ulcerative colitis (UC). Diagnosis of histologic remission currently relies on biopsy; during this process, patients are at risk for bleeding, infection, and post-biopsy fibrosis. In addition, histologic response scoring is complex and time-consuming, and there is heterogeneity among pathologists. Endocytoscopy (EC) is a novel ultra-high magnification endoscopic technique that can provide excellent in vivo assessment of glands. Based on the EC technique, we propose a neural network model that can assess histological disease activity in UC using EC images to address the above issues. The experiment results demonstrate that the proposed method can assist patients in precise treatment and prognostic assessment. METHODS: We construct a neural network model for UC evaluation. A total of 5105 images of 154 intestinal segments from 87 patients undergoing EC treatment at a center in China between March 2022 and March 2023 are scored according to the Geboes score. Subsequently, 103 intestinal segments are used as the training set, 16 intestinal segments are used as the validation set for neural network training, and the remaining 35 intestinal segments are used as the test set to measure the model performance together with the validation set. RESULTS: By treating HR as a negative category and histologic activity as a positive category, the proposed neural network model can achieve an accuracy of 0.9, a specificity of 0.95, a sensitivity of 0.75, and an area under the curve (AUC) of 0.81. CONCLUSION: We develop a specific neural network model that can distinguish histologic remission/activity in EC images of UC, which helps to accelerate clinical histological diagnosis. keywords: ulcerative colitis; Endocytoscopy; Geboes score; neural network.

Abstract: There is a growing interest in using pose estimation algorithms for video-based assessment of Bradykinesia in Parkinson's Disease (PD) to facilitate remote disease assessment and monitoring. However, the accuracy of pose estimation algorithms in videos from video streaming services during Telehealth appointments has not been studied. In this study, we used seven off-the-shelf hand pose estimation models to estimate the movement of the thumb and index fingers in videos of the finger-tapping (FT) test recorded from Healthy Controls (HC) and participants with PD and under two different conditions: streaming (videos recorded during a live Zoom meeting) and on-device (videos recorded locally with high-quality cameras). The accuracy and reliability of the models were estimated by comparing the models' output with manual results. Three of the seven models demonstrated good accuracy for on-device recordings, and the accuracy decreased significantly for streaming recordings. We observed a negative correlation between movement speed and the model's accuracy for the streaming recordings. Additionally, we evaluated the reliability of ten movement features related to bradykinesia extracted from video recordings of PD patients performing the FT test. While most of the features demonstrated excellent reliability for on-device recordings, most of the features demonstrated poor to moderate reliability for streaming recordings. Our findings highlight the limitations of pose estimation algorithms when applied to video recordings obtained during Telehealth visits, and demonstrate that on-device recordings can be used for automatic video-assessment of bradykinesia in PD.

Abstract: 360$^\circ$ panoramas are extensively utilized as environmental light sources in computer graphics. However, capturing a 360$^\circ$ $\times$ 180$^\circ$ panorama poses challenges due to the necessity of specialized and costly equipment, and additional human resources. Prior studies develop various learning-based generative methods to synthesize panoramas from a single Narrow Field-of-View (NFoV) image, but they are limited in alterable input patterns, generation quality, and controllability. To address these issues, we propose a novel pipeline called PanoDiff, which efficiently generates complete 360$^\circ$ panoramas using one or more unregistered NFoV images captured from arbitrary angles. Our approach has two primary components to overcome the limitations. Firstly, a two-stage angle prediction module to handle various numbers of NFoV inputs. Secondly, a novel latent diffusion-based panorama generation model uses incomplete panorama and text prompts as control signals and utilizes several geometric augmentation schemes to ensure geometric properties in generated panoramas. Experiments show that PanoDiff achieves state-of-the-art panoramic generation quality and high controllability, making it suitable for applications such as content editing.

Abstract: Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving 50.7% APvideo^50 , surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by 15.9% on YouTubeVIS-2019 in terms of APvideo.

Abstract: Dense 3D reconstruction and ego-motion estimation are key challenges in autonomous driving and robotics. Compared to the complex, multi-modal systems deployed today, multi-camera systems provide a simpler, low-cost alternative. However, camera-based 3D reconstruction of complex dynamic scenes has proven extremely difficult, as existing solutions often produce incomplete or incoherent results. We propose R3D3, a multi-camera system for dense 3D reconstruction and ego-motion estimation. Our approach iterates between geometric estimation that exploits spatial-temporal information from multiple cameras, and monocular depth refinement. We integrate multi-camera feature correlation and dense bundle adjustment operators that yield robust geometric depth and pose estimates. To improve reconstruction where geometric depth is unreliable, e.g. for moving objects or low-textured regions, we introduce learnable scene priors via a depth refinement network. We show that this design enables a dense, consistent 3D reconstruction of challenging, dynamic outdoor environments. Consequently, we achieve state-of-the-art dense depth prediction on the DDAD and NuScenes benchmarks.

Abstract: In panorama understanding, the widely used equirectangular projection (ERP) entails boundary discontinuity and spatial distortion. It severely deteriorates the conventional CNNs and vision Transformers on panoramas. In this paper, we propose a simple yet effective architecture named PanoSwin to learn panorama representations with ERP. To deal with the challenges brought by equirectangular projection, we explore a pano-style shift windowing scheme and novel pitch attention to address the boundary discontinuity and the spatial distortion, respectively. Besides, based on spherical distance and Cartesian coordinates, we adapt absolute positional embeddings and relative positional biases for panoramas to enhance panoramic geometry information. Realizing that planar image understanding might share some common knowledge with panorama understanding, we devise a novel two-stage learning framework to facilitate knowledge transfer from the planar images to panoramas. We conduct experiments against the state-of-the-art on various panoramic tasks, i.e., panoramic object detection, panoramic classification, and panoramic layout estimation. The experimental results demonstrate the effectiveness of PanoSwin in panorama understanding.

Abstract: Fast, reliable shape reconstruction is an essential ingredient in many computer vision applications. Neural Radiance Fields demonstrated that photorealistic novel view synthesis is within reach, but was gated by performance requirements for fast reconstruction of real scenes and objects. Several recent approaches have built on alternative shape representations, in particular, 3D Gaussians. We develop extensions to these renderers, such as integrating differentiable optical flow, exporting watertight meshes and rendering per-ray normals. Additionally, we show how two of the recent methods are interoperable with each other. These reconstructions are quick, robust, and easily performed on GPU or CPU. For code and visual examples, see https://leonidk.github.io/fmb-plus

Abstract: We present a method to generate full-body selfies -- photos that you take of yourself, but capturing your whole body as if someone else took the photo of you from a few feet away. Our approach takes as input a pre-captured video of your body, a target pose photo, and a selfie + background pair for each location. We introduce a novel diffusion-based approach to combine all of this information into high quality, well-composed photos of you with the desired pose and background.

Abstract: Composed Image Retrieval (CoIR) has recently gained popularity as a task that considers both text and image queries together, to search for relevant images in a database. Most CoIR approaches require manually annotated datasets, comprising image-text-image triplets, where the text describes a modification from the query image to the target image. However, manual curation of CoIR triplets is expensive and prevents scalability. In this work, we instead propose a scalable automatic dataset creation methodology that generates triplets given video-caption pairs, while also expanding the scope of the task to include composed video retrieval (CoVR). To this end, we mine paired videos with a similar caption from a large database, and leverage a large language model to generate the corresponding modification text. Applying this methodology to the extensive WebVid2M collection, we automatically construct our WebVid-CoVR dataset, resulting in 1.6 million triplets. Moreover, we introduce a new benchmark for CoVR with a manually annotated evaluation set, along with baseline results. Our experiments further demonstrate that training a CoVR model on our dataset effectively transfers to CoIR, leading to improved state-of-the-art performance in the zero-shot setup on both the CIRR and FashionIQ benchmarks. Our code, datasets, and models are publicly available at https://imagine.enpc.fr/~ventural/covr.

Abstract: In this report, we present MagicEdit, a surprisingly simple yet effective solution to the text-guided video editing task. We found that high-fidelity and temporally coherent video-to-video translation can be achieved by explicitly disentangling the learning of content, structure and motion signals during training. This is in contradict to most existing methods which attempt to jointly model both the appearance and temporal representation within a single framework, which we argue, would lead to degradation in per-frame quality. Despite its simplicity, we show that MagicEdit supports various downstream video editing tasks, including video stylization, local editing, video-MagicMix and video outpainting.

Abstract: Visual similarities discovery (VSD) is an important task with broad e-commerce applications. Given an image of a certain object, the goal of VSD is to retrieve images of different objects with high perceptual visual similarity. Although being a highly addressed problem, the evaluation of proposed methods for VSD is often based on a proxy of an identification-retrieval task, evaluating the ability of a model to retrieve different images of the same object. We posit that evaluating VSD methods based on identification tasks is limited, and faithful evaluation must rely on expert annotations. In this paper, we introduce the first large-scale fashion visual similarity benchmark dataset, consisting of more than 110K expert-annotated image pairs. Besides this major contribution, we share insight from the challenges we faced while curating this dataset. Based on these insights, we propose a novel and efficient labeling procedure that can be applied to any dataset. Our analysis examines its limitations and inductive biases, and based on these findings, we propose metrics to mitigate those limitations. Though our primary focus lies on visual similarity, the methodologies we present have broader applications for discovering and evaluating perceptual similarity across various domains.