In this paper, we argue that iterative computation with diffusion models offers a powerful paradigm for not only generation but also visual perception tasks. We unify tasks such as depth estimation, optical flow, and amodal segmentation under the framework of image-to-image translation, and show how diffusion models benefit from scaling training and test-time compute for these perceptual tasks. Through a careful analysis of these scaling properties, we formulate compute-optimal training and inference recipes to scale diffusion models for visual perception tasks. Our models achieve competitive performance to state-of-the-art methods using significantly less data and compute. To access our code and models, see https://scaling-diffusion-perception.github.io .
Rapid ice recession in the Arctic Ocean, with predictions of ice-free summers by 2060, opens new maritime routes but requires reliable navigation solutions. Current approaches rely heavily on subjective expert judgment, underscoring the need for automated, data-driven solutions. This study leverages machine learning to assess ice conditions using ship-borne optical data, introducing a finely annotated dataset of 946 images, and a semi-manual, region-based annotation technique. The proposed video segmentation model, UPerFlow, advances the SegFlow architecture by incorporating a six-channel ResNet encoder, two UPerNet-based segmentation decoders for each image, PWCNet as the optical flow encoder, and cross-connections that integrate bi-directional flow features without loss of latent information. The proposed architecture outperforms baseline image segmentation networks by an average 38% in occluded regions, demonstrating the robustness of video segmentation in addressing challenging Arctic conditions.
Surgical instrument segmentation (SIS) is pivotal for robotic-assisted minimally invasive surgery, assisting surgeons by identifying surgical instruments in endoscopic video frames. Recent unsupervised surgical instrument segmentation (USIS) methods primarily rely on pseudo-labels derived from low-level features such as color and optical flow, but these methods show limited effectiveness and generalizability in complex and unseen endoscopic scenarios. In this work, we propose a label-free unsupervised model featuring a novel module named Multi-View Normalized Cutter (m-NCutter). Different from previous USIS works, our model is trained using a graph-cutting loss function that leverages patch affinities for supervision, eliminating the need for pseudo-labels. The framework adaptively determines which affinities from which levels should be prioritized. Therefore, the low- and high-level features and their affinities are effectively integrated to train a label-free unsupervised model, showing superior effectiveness and generalization ability. We conduct comprehensive experiments across multiple SIS datasets to validate our approach's state-of-the-art (SOTA) performance, robustness, and exceptional potential as a pre-trained model. Our code is released at https://github.com/MingyuShengSMY/AMNCutter.
As violent crimes continue to happen, it becomes necessary to have security cameras that can rapidly identify moments of violence with excellent accuracy. The purpose of this study is to identify how many frames should be analyzed at a time in order to optimize a violence detection model's accuracy as a parameter of the depth of a 3D convolutional network. Previous violence classification models have been created, but their application to live footage may be flawed. In this project, a convolutional neural network was created to analyze optical flow frames of each video. The number of frames analyzed at a time would vary with one, two, three, ten, and twenty frames, and each model would be trained for 20 epochs. The greatest validation accuracy was 94.87% and occurred with the model that analyzed three frames at a time. This means that machine learning models to detect violence may function better when analyzing three frames at a time for this dataset. The methodology used to identify the optimal number of frames to analyze at a time could be used in other applications of video classification, especially those of complex or abstract actions, such as violence.
Humans excel at detecting and segmenting moving objects according to the Gestalt principle of "common fate". Remarkably, previous works have shown that human perception generalizes this principle in a zero-shot fashion to unseen textures or random dots. In this work, we seek to better understand the computational basis for this capability by evaluating a broad range of optical flow models and a neuroscience inspired motion energy model for zero-shot figure-ground segmentation of random dot stimuli. Specifically, we use the extensively validated motion energy model proposed by Simoncelli and Heeger in 1998 which is fitted to neural recordings in cortex area MT. We find that a cross section of 40 deep optical flow models trained on different datasets struggle to estimate motion patterns in random dot videos, resulting in poor figure-ground segmentation performance. Conversely, the neuroscience-inspired model significantly outperforms all optical flow models on this task. For a direct comparison to human perception, we conduct a psychophysical study using a shape identification task as a proxy to measure human segmentation performance. All state-of-the-art optical flow models fall short of human performance, but only the motion energy model matches human capability. This neuroscience-inspired model successfully addresses the lack of human-like zero-shot generalization to random dot stimuli in current computer vision models, and thus establishes a compelling link between the Gestalt psychology of human object perception and cortical motion processing in the brain. Code, models and datasets are available at https://github.com/mtangemann/motion_energy_segmentation
Medical video generation models are expected to have a profound impact on the healthcare industry, including but not limited to medical education and training, surgical planning, and simulation. Current video diffusion models typically build on image diffusion architecture by incorporating temporal operations (such as 3D convolution and temporal attention). Although this approach is effective, its oversimplification limits spatio-temporal performance and consumes substantial computational resources. To counter this, we propose Medical Simulation Video Generator (MedSora), which incorporates three key elements: i) a video diffusion framework integrates the advantages of attention and Mamba, balancing low computational load with high-quality video generation, ii) an optical flow representation alignment method that implicitly enhances attention to inter-frame pixels, and iii) a video variational autoencoder (VAE) with frequency compensation addresses the information loss of medical features that occurs when transforming pixel space into latent features and then back to pixel frames. Extensive experiments and applications demonstrate that MedSora exhibits superior visual quality in generating medical videos, outperforming the most advanced baseline methods. Further results and code are available at https://wongzbb.github.io/MedSora
Cloud cover plays a critical role in weather prediction and impacts several sectors, including agriculture, solar power generation, and aviation. Despite advancements in numerical weather prediction (NWP) models, forecasting total cloud cover remains challenging due to the small-scale nature of cloud formation processes. In this study, we introduce CloudCast, a convolutional neural network (CNN) based on the U-Net architecture, designed to predict total cloud cover (TCC) up to five hours ahead. Trained on five years of satellite data, CloudCast significantly outperforms traditional NWP models and optical flow methods. Compared to a reference NWP model, CloudCast achieves a 24% lower mean absolute error and reduces multi-category prediction errors by 46%. The model demonstrates strong performance, particularly in capturing the large-scale structure of cloud cover in the first few forecast hours, though later predictions are subject to blurring and underestimation of cloud formation. An ablation study identified the optimal input features and loss functions, with MAE-based models performing the best. CloudCast has been integrated into the Finnish Meteorological Institute's operational nowcasting system, where it improves cloud cover forecasts used by public and private sector clients. While CloudCast is limited by a relatively short skillful lead time of about three hours, future work aims to extend this through more complex network architectures and higher-resolution data. CloudCast code is available at https://github.com/fmidev/cloudcast.
We introduce motion graph, a novel approach to the video prediction problem, which predicts future video frames from limited past data. The motion graph transforms patches of video frames into interconnected graph nodes, to comprehensively describe the spatial-temporal relationships among them. This representation overcomes the limitations of existing motion representations such as image differences, optical flow, and motion matrix that either fall short in capturing complex motion patterns or suffer from excessive memory consumption. We further present a video prediction pipeline empowered by motion graph, exhibiting substantial performance improvements and cost reductions. Experiments on various datasets, including UCF Sports, KITTI and Cityscapes, highlight the strong representative ability of motion graph. Especially on UCF Sports, our method matches and outperforms the SOTA methods with a significant reduction in model size by 78% and a substantial decrease in GPU memory utilization by 47%.
Reconstructing controllable Gaussian splats from monocular video is a challenging task due to its inherently insufficient constraints. Widely adopted approaches supervise complex interactions with additional masks and control signal annotations, limiting their real-world applications. In this paper, we propose an annotation guidance-free method, dubbed FreeGaussian, that mathematically derives dynamic Gaussian motion from optical flow and camera motion using novel dynamic Gaussian constraints. By establishing a connection between 2D flows and 3D Gaussian dynamic control, our method enables self-supervised optimization and continuity of dynamic Gaussian motions from flow priors. Furthermore, we introduce a 3D spherical vector controlling scheme, which represents the state with a 3D Gaussian trajectory, thereby eliminating the need for complex 1D control signal calculations and simplifying controllable Gaussian modeling. Quantitative and qualitative evaluations on extensive experiments demonstrate the state-of-the-art visual performance and control capability of our method. Project page: https://freegaussian.github.io.
In this work, we extract the optical flow field corresponding to moving objects from an image sequence of a scene impacted by atmospheric turbulence \emph{and} captured from a moving camera. Our procedure first computes the optical flow field and creates a motion model to compensate for the flow field induced by camera motion. After subtracting the motion model from the optical flow, we proceed with our previous work, Gilles et al~\cite{gilles2018detection}, where a spatial-temporal cartoon+texture inspired decomposition is performed on the motion-compensated flow field in order to separate flows corresponding to atmospheric turbulence and object motion. Finally, the geometric component is processed with the detection and tracking method and is compared against a ground truth. All of the sequences and code used in this work are open source and are available by contacting the authors.
Recent advances in event-based vision suggest that these systems complement traditional cameras by providing continuous observation without frame rate limitations and a high dynamic range, making them well-suited for correspondence tasks such as optical flow and point tracking. However, there is still a lack of comprehensive benchmarks for correspondence tasks that include both event data and images. To address this gap, we propose BlinkVision, a large-scale and diverse benchmark with multiple modalities and dense correspondence annotations. BlinkVision offers several valuable features: 1) Rich modalities: It includes both event data and RGB images. 2) Extensive annotations: It provides dense per-pixel annotations covering optical flow, scene flow, and point tracking. 3) Large vocabulary: It contains 410 everyday categories, sharing common classes with popular 2D and 3D datasets like LVIS and ShapeNet. 4) Naturalistic: It delivers photorealistic data and covers various naturalistic factors, such as camera shake and deformation. BlinkVision enables extensive benchmarks on three types of correspondence tasks (optical flow, point tracking, and scene flow estimation) for both image-based and event-based methods, offering new observations, practices, and insights for future research. The benchmark website is https://www.blinkvision.net/.
The sequential execution of actions and their hierarchical structure consisting of different levels of abstraction, provide features that remain unexplored in the task of action recognition. In this study, we present a novel approach to improve action recognition by exploiting the hierarchical organization of actions and by incorporating contextualized textual information, including location and prior actions to reflect the sequential context. To achieve this goal, we introduce a novel transformer architecture tailored for action recognition that utilizes both visual and textual features. Visual features are obtained from RGB and optical flow data, while text embeddings represent contextual information. Furthermore, we define a joint loss function to simultaneously train the model for both coarse and fine-grained action recognition, thereby exploiting the hierarchical nature of actions. To demonstrate the effectiveness of our method, we extend the Toyota Smarthome Untrimmed (TSU) dataset to introduce action hierarchies, introducing the Hierarchical TSU dataset. We also conduct an ablation study to assess the impact of different methods for integrating contextual and hierarchical data on action recognition performance. Results show that the proposed approach outperforms pre-trained SOTA methods when trained with the same hyperparameters. Moreover, they also show a 17.12% improvement in top-1 accuracy over the equivalent fine-grained RGB version when using ground-truth contextual information, and a 5.33% improvement when contextual information is obtained from actual predictions.
Intelligent processing and analysis of satellite video has become one of the research hotspots in the representation of remote sensing, and satellite video super-resolution (SVSR) is an important research direction which can improve the image quality of satellite video. However, existing approaches for SVSR often underutilize a notable advantage inherent to satellite video: the presence of extensive sequential imagery capturing a consistent scene. Presently, the majority of SVSR methods merely harness a limited number of adjacent frames for enhancing the resolution of individual frames, thus resulting in suboptimal information utilization. In response, we introduce the Recurrent Aggregation Network for Satellite Video Super-Resolution (RASVSR). This innovative framework leverages a bidirectional recurrent neural network to propagate extracted features from each frame across the entire video sequence. It relies on an alignment method based on optical flow and deformable convolution (DCN) to realize the alignment of the features, and a Temporal Feature Fusion Module (TFF) to realize effective feature fusion over time. Notably, our research underscores the positive influence of employing lengthier image sequences in SVSR. In the context of RASVSR, with better alignment and fusion, we make the perceptual field of each frame spanning 100 frames of the video, thus acquiring richer information, and information between different images can be complementary. This strategic approach culminates in superior performance compared to alternative methods, as evidenced by a noteworthy 1.15 dB improvement in PSNR, with very few parameters.
In the domain of computer vision, optical flow stands as a cornerstone for unraveling dynamic visual scenes. However, the challenge of accurately estimating optical flow under conditions of large nonlinear motion patterns remains an open question. The image flow constraint is vulnerable to substantial displacements, and rapid spatial transformations. Inaccurate approximations inherent in numerical differentiation techniques can further amplify such intricacies. In response, this research proposes an innovative algorithm for optical flow computation, utilizing the higher precision of second-order Taylor series approximation within the differential estimation framework. By embracing this mathematical underpinning, the research seeks to extract more information about the behavior of the function under complex real-world scenarios and estimate the motion of areas with a lack of texture. An impressive showcase of the algorithm's capabilities emerges through its performance on renowned optical flow benchmarks such as KITTI (2015) and Middlebury. The average endpoint error (AEE), which computes the Euclidian distance between the calculated flow field and the ground truth flow field, stands notably diminished, validating the effectiveness of the algorithm in handling complex motion patterns.
A significant challenge facing current optical flow and stereo methods is the difficulty in generalizing them well to the real world. This is mainly due to the high costs required to produce datasets, and the limitations of existing self-supervised methods on fuzzy results and complex model training problems. To address the above challenges, we propose a unified self-supervised generalization framework for optical flow and stereo tasks: Self-Assessed Generation (SAG). Unlike previous self-supervised methods, SAG is data-driven, using advanced reconstruction techniques to construct a reconstruction field from RGB images and generate datasets based on it. Afterward, we quantified the confidence level of the generated results from multiple perspectives, such as reconstruction field distribution, geometric consistency, and structural similarity, to eliminate inevitable defects in the generation process. We also designed a 3D flight foreground automatic rendering pipeline in SAG to encourage the network to learn occlusion and motion foreground. Experimentally, because SAG does not involve changes to methods or loss functions, it can directly self-supervised train the state-of-the-art deep networks, greatly improving the generalization performance of self-supervised methods on current mainstream optical flow and stereo-matching datasets. Compared to previous training modes, SAG is more generalized, cost-effective, and accurate.
Dynamic scene reconstruction is a long-term challenge in the field of 3D vision. Recently, the emergence of 3D Gaussian Splatting has provided new insights into this problem. Although subsequent efforts rapidly extend static 3D Gaussian to dynamic scenes, they often lack explicit constraints on object motion, leading to optimization difficulties and performance degradation. To address the above issues, we propose a novel deformable 3D Gaussian splatting framework called MotionGS, which explores explicit motion priors to guide the deformation of 3D Gaussians. Specifically, we first introduce an optical flow decoupling module that decouples optical flow into camera flow and motion flow, corresponding to camera movement and object motion respectively. Then the motion flow can effectively constrain the deformation of 3D Gaussians, thus simulating the motion of dynamic objects. Additionally, a camera pose refinement module is proposed to alternately optimize 3D Gaussians and camera poses, mitigating the impact of inaccurate camera poses. Extensive experiments in the monocular dynamic scenes validate that MotionGS surpasses state-of-the-art methods and exhibits significant superiority in both qualitative and quantitative results. Project page: https://ruijiezhu94.github.io/MotionGS_page
Lung radiotherapy treatment systems are subject to a latency that leads to uncertainty in the estimated tumor location and high irradiation of healthy tissue. This work addresses future frame prediction in chest dynamic MRI sequences to compensate for that delay using RNNs trained with online learning algorithms. The latter enable networks to mitigate irregular movements, as they update synaptic weights with each new training example. Experiments were conducted using four publicly available 2D thoracic cine-MRI sequences. PCA decomposes the time-varying deformation vector field (DVF), computed with the Lucas-Kanade optical flow algorithm, into static deformation fields and low-dimensional time-dependent weights. We compare various algorithms to forecast the latter: linear regression, least mean squares (LMS), and RNNs trained with real-time recurrent learning (RTRL), unbiased online recurrent optimization, decoupled neural interfaces and sparse 1-step approximation (SnAp-1). That enables estimating the future DVFs and, in turn, the next frames by warping the initial image. Linear regression led to the lowest mean DVF error at a horizon h = 0.32s (the time interval in advance for which the prediction is made), equal to 1.30mm, followed by SnAp-1 and RTRL, whose error increased from 1.37mm to 1.44mm as h increased from 0.62s to 2.20s. Similarly, the structural similarity index measure (SSIM) of LMS decreased from 0.904 to 0.898 as h increased from 0.31s to 1.57s and was the highest among the algorithms compared for the latter horizons. SnAp-1 attained the highest SSIM for h $\geq$ 1.88s, with values of less than 0.898. The predicted images look similar to the original ones, and the highest errors occurred at challenging areas such as the diaphragm boundary at the end-of-inhale phase, where motion variability is more prominent, and regions where out-of-plane motion was more prevalent.
We propose a novel optical flow based approach to enhance the axial resolution of anisotropic 3D EM volumes to achieve isotropic 3D reconstruction. Assuming spatial continuity of 3D biological structures in well aligned EM volumes, we reasoned that optical flow estimation techniques, often applied for temporal resolution enhancement in videos, can be utilized. Pixel level motion is estimated between neighboring 2D slices along z, using spatial gradient flow estimates to interpolate and generate new 2D slices resulting in isotropic voxels. We leverage recent state-of-the-art learning methods for video frame interpolation and transfer learning techniques, and demonstrate the success of our approach on publicly available ultrastructure EM volumes.
Smart focal-plane and in-chip image processing has emerged as a crucial technology for vision-enabled embedded systems with energy efficiency and privacy. However, the lack of special datasets providing examples of the data that these neuromorphic sensors compute to convey visual information has hindered the adoption of these promising technologies. Neuromorphic imager variants, including event-based sensors, produce various representations such as streams of pixel addresses representing time and locations of intensity changes in the focal plane, temporal-difference data, data sifted/thresholded by temporal differences, image data after applying spatial transformations, optical flow data, and/or statistical representations. To address the critical barrier to entry, we provide an annotated, temporal-threshold-based vision dataset specifically designed for face detection tasks derived from the same videos used for Aff-Wild2. By offering multiple threshold levels (e.g., 4, 8, 12, and 16), this dataset allows for comprehensive evaluation and optimization of state-of-the-art neural architectures under varying conditions and settings compared to traditional methods. The accompanying tool flow for generating event data from raw videos further enhances accessibility and usability. We anticipate that this resource will significantly support the development of robust vision systems based on smart sensors that can process based on temporal-difference thresholds, enabling more accurate and efficient object detection and localization and ultimately promoting the broader adoption of low-power, neuromorphic imaging technologies. To support further research, we publicly released the dataset at \url{https://dx.doi.org/10.21227/bw2e-dj78}.
We present a method to reconstruct time-consistent human body models from monocular videos, focusing on extremely loose clothing or handheld object interactions. Prior work in human reconstruction is either limited to tight clothing with no object interactions, or requires calibrated multi-view captures or personalized template scans which are costly to collect at scale. Our key insight for high-quality yet flexible reconstruction is the careful combination of generic human priors about articulated body shape (learned from large-scale training data) with video-specific articulated "bag-of-bones" deformation (fit to a single video via test-time optimization). We accomplish this by learning a neural implicit model that disentangles body versus clothing deformations as separate motion model layers. To capture subtle geometry of clothing, we leverage image-based priors such as human body pose, surface normals, and optical flow during optimization. The resulting neural fields can be extracted into time-consistent meshes, or further optimized as explicit 3D Gaussians for high-fidelity interactive rendering. On datasets with highly challenging clothing deformations and object interactions, DressRecon yields higher-fidelity 3D reconstructions than prior art. Project page: https://jefftan969.github.io/dressrecon/