The 3D contrastive learning paradigm has demonstrated remarkable performance in downstream tasks through pretraining on point cloud data. Recent advances involve additional 2D image priors associated with 3D point clouds for further improvement. Nonetheless, these existing frameworks are constrained by the restricted range of available point cloud datasets, primarily due to the high costs of obtaining point cloud data. To this end, we propose SimC3D, a simple but effective 3D contrastive learning framework, for the first time, pretraining 3D backbones from pure RGB image data. SimC3D performs contrastive 3D pretraining with three appealing properties. (1) Pure image data: SimC3D simplifies the dependency of costly 3D point clouds and pretrains 3D backbones using solely RBG images. By employing depth estimation and suitable data processing, the monocular synthesized point cloud shows great potential for 3D pretraining. (2) Simple framework: Traditional multi-modal frameworks facilitate 3D pretraining with 2D priors by utilizing an additional 2D backbone, thereby increasing computational expense. In this paper, we empirically demonstrate that the primary benefit of the 2D modality stems from the incorporation of locality information. Inspired by this insightful observation, SimC3D directly employs 2D positional embeddings as a stronger contrastive objective, eliminating the necessity for 2D backbones and leading to considerable performance improvements. (3) Strong performance: SimC3D outperforms previous approaches that leverage ground-truth point cloud data for pretraining in various downstream tasks. Furthermore, the performance of SimC3D can be further enhanced by combining multiple image datasets, showcasing its significant potential for scalability. The code will be available at https://github.com/Dongjiahua/SimC3D.
In this paper, we present PanoDreamer, a novel method for producing a coherent 360{\deg} 3D scene from a single input image. Unlike existing methods that generate the scene sequentially, we frame the problem as single-image panorama and depth estimation. Once the coherent panoramic image and its corresponding depth are obtained, the scene can be reconstructed by inpainting the small occluded regions and projecting them into 3D space. Our key contribution is formulating single-image panorama and depth estimation as two optimization tasks and introducing alternating minimization strategies to effectively solve their objectives. We demonstrate that our approach outperforms existing techniques in single-image 360{\deg} 3D scene reconstruction in terms of consistency and overall quality.
The choice of data representation is a key factor in the success of deep learning in geometric tasks. For instance, DUSt3R has recently introduced the concept of viewpoint-invariant point maps, generalizing depth prediction, and showing that one can reduce all the key problems in the 3D reconstruction of static scenes to predicting such point maps. In this paper, we develop an analogous concept for a very different problem, namely, the reconstruction of the 3D shape and pose of deformable objects. To this end, we introduce the Dual Point Maps (DualPM), where a pair of point maps is extracted from the same image, one associating pixels to their 3D locations on the object, and the other to a canonical version of the object at rest pose. We also extend point maps to amodal reconstruction, seeing through self-occlusions to obtain the complete shape of the object. We show that 3D reconstruction and 3D pose estimation reduce to the prediction of the DualPMs. We demonstrate empirically that this representation is a good target for a deep network to predict; specifically, we consider modeling horses, showing that DualPMs can be trained purely on 3D synthetic data, consisting of a single model of a horse, while generalizing very well to real images. With this, we improve by a large margin previous methods for the 3D analysis and reconstruction of this type of objects.
We present a system that allows for accurate, fast, and robust estimation of camera parameters and depth maps from casual monocular videos of dynamic scenes. Most conventional structure from motion and monocular SLAM techniques assume input videos that feature predominantly static scenes with large amounts of parallax. Such methods tend to produce erroneous estimates in the absence of these conditions. Recent neural network-based approaches attempt to overcome these challenges; however, such methods are either computationally expensive or brittle when run on dynamic videos with uncontrolled camera motion or unknown field of view. We demonstrate the surprising effectiveness of a deep visual SLAM framework: with careful modifications to its training and inference schemes, this system can scale to real-world videos of complex dynamic scenes with unconstrained camera paths, including videos with little camera parallax. Extensive experiments on both synthetic and real videos demonstrate that our system is significantly more accurate and robust at camera pose and depth estimation when compared with prior and concurrent work, with faster or comparable running times. See interactive results on our project page: https://mega-sam.github.io/
In image-assisted minimally invasive surgeries (MIS), understanding surgical scenes is vital for real-time feedback to surgeons, skill evaluation, and improving outcomes through collaborative human-robot procedures. Within this context, the challenge lies in accurately detecting, segmenting, and estimating the depth of surgical scenes depicted in high-resolution images, while simultaneously reconstructing the scene in 3D and providing segmentation of surgical instruments along with detection labels for each instrument. To address this challenge, a novel Multi-Task Learning (MTL) network is proposed for performing these tasks concurrently. A key aspect of this approach involves overcoming the optimization hurdles associated with handling multiple tasks concurrently by integrating a Adversarial Weight Update into the MTL framework, the proposed MTL model achieves 3D reconstruction through the integration of segmentation, depth estimation, and object detection, thereby enhancing the understanding of surgical scenes, which marks a significant advancement compared to existing studies that lack 3D capabilities. Comprehensive experiments on the EndoVis2018 benchmark dataset underscore the adeptness of the model in efficiently addressing all three tasks, demonstrating the efficacy of the proposed techniques.
This paper investigates the research task of reconstructing the 3D clothed human body from a monocular image. Due to the inherent ambiguity of single-view input, existing approaches leverage pre-trained SMPL(-X) estimation models or generative models to provide auxiliary information for human reconstruction. However, these methods capture only the general human body geometry and overlook specific geometric details, leading to inaccurate skeleton reconstruction, incorrect joint positions, and unclear cloth wrinkles. In response to these issues, we propose a multi-level geometry learning framework. Technically, we design three key components: skeleton-level enhancement, joint-level augmentation, and wrinkle-level refinement modules. Specifically, we effectively integrate the projected 3D Fourier features into a Gaussian reconstruction model, introduce perturbations to improve joint depth estimation during training, and refine the human coarse wrinkles by resembling the de-noising process of diffusion model. Extensive quantitative and qualitative experiments on two out-of-distribution test sets show the superior performance of our approach compared to state-of-the-art (SOTA) methods.
Recent developments in monocular depth estimation methods enable high-quality depth estimation of single-view images but fail to estimate consistent video depth across different frames. Recent works address this problem by applying a video diffusion model to generate video depth conditioned on the input video, which is training-expensive and can only produce scale-invariant depth values without camera poses. In this paper, we propose a novel video-depth estimation method called Align3R to estimate temporal consistent depth maps for a dynamic video. Our key idea is to utilize the recent DUSt3R model to align estimated monocular depth maps of different timesteps. First, we fine-tune the DUSt3R model with additional estimated monocular depth as inputs for the dynamic scenes. Then, we apply optimization to reconstruct both depth maps and camera poses. Extensive experiments demonstrate that Align3R estimates consistent video depth and camera poses for a monocular video with superior performance than baseline methods.
This paper presents a dense depth estimation approach from light-field (LF) images that is able to compensate for strong rolling shutter (RS) effects. Our method estimates RS compensated views and dense RS compensated disparity maps. We present a two-stage method based on a 2D Gaussians Splatting that allows for a ``render and compare" strategy with a point cloud formulation. In the first stage, a subset of sub-aperture images is used to estimate an RS agnostic 3D shape that is related to the scene target shape ``up to a motion". In the second stage, the deformation of the 3D shape is computed by estimating an admissible camera motion. We demonstrate the effectiveness and advantages of this approach through several experiments conducted for different scenes and types of motions. Due to lack of suitable datasets for evaluation, we also present a new carefully designed synthetic dataset of RS LF images. The source code, trained models and dataset will be made publicly available at: https://github.com/ICB-Vision-AI/DenseRSLF
Multimodal language models (MLMs) still face challenges in fundamental visual perception tasks where specialized models excel. Tasks requiring reasoning about 3D structures benefit from depth estimation, and reasoning about 2D object instances benefits from object detection. Yet, MLMs can not produce intermediate depth or boxes to reason over. Finetuning MLMs on relevant data doesn't generalize well and outsourcing computation to specialized vision tools is too compute-intensive and memory-inefficient. To address this, we introduce Perception Tokens, intrinsic image representations designed to assist reasoning tasks where language is insufficient. Perception tokens act as auxiliary reasoning tokens, akin to chain-of-thought prompts in language models. For example, in a depth-related task, an MLM augmented with perception tokens can reason by generating a depth map as tokens, enabling it to solve the problem effectively. We propose AURORA, a training method that augments MLMs with perception tokens for improved reasoning over visual inputs. AURORA leverages a VQVAE to transform intermediate image representations, such as depth maps into a tokenized format and bounding box tokens, which is then used in a multi-task training framework. AURORA achieves notable improvements across counting benchmarks: +10.8% on BLINK, +11.3% on CVBench, and +8.3% on SEED-Bench, outperforming finetuning approaches in generalization across datasets. It also improves on relative depth: over +6% on BLINK. With perception tokens, AURORA expands the scope of MLMs beyond language-based reasoning, paving the way for more effective visual reasoning capabilities.
Video monocular depth estimation is essential for applications such as autonomous driving, AR/VR, and robotics. Recent transformer-based single-image monocular depth estimation models perform well on single images but struggle with depth consistency across video frames. Traditional methods aim to improve temporal consistency using multi-frame temporal modules or prior information like optical flow and camera parameters. However, these approaches face issues such as high memory use, reduced performance with dynamic or irregular motion, and limited motion understanding. We propose STATIC, a novel model that independently learns temporal consistency in static and dynamic area without additional information. A difference mask from surface normals identifies static and dynamic area by measuring directional variance. For static area, the Masked Static (MS) module enhances temporal consistency by focusing on stable regions. For dynamic area, the Surface Normal Similarity (SNS) module aligns areas and enhances temporal consistency by measuring feature similarity between frames. A final refinement integrates the independently learned static and dynamic area, enabling STATIC to achieve temporal consistency across the entire sequence. Our method achieves state-of-the-art video depth estimation on the KITTI and NYUv2 datasets without additional information.
Monocular Depth Estimation (MDE) is essential for applications like 3D scene reconstruction, autonomous navigation, and AI content creation. However, robust MDE remains challenging due to noisy real-world data and distribution gaps in synthetic datasets. Existing methods often struggle with low efficiency, reduced accuracy, and lack of detail. To address this, we propose an efficient approach for leveraging diffusion priors and introduce FiffDepth, a framework that transforms diffusion-based image generators into a feedforward architecture for detailed depth estimation. By preserving key generative features and integrating the strong generalization capabilities of models like dinov2, FiffDepth achieves enhanced accuracy, stability, and fine-grained detail, offering a significant improvement in MDE performance across diverse real-world scenarios.
Metric depth prediction from monocular videos suffers from bad generalization between datasets and requires supervised depth data for scale-correct training. Self-supervised training using multi-view reconstruction can benefit from large scale natural videos but not provide correct scale, limiting its benefits. Recently, reflecting audible Echoes off objects is investigated for improved depth prediction and was shown to be sufficient to reconstruct objects at scale even without a visual signal. Because Echoes travel at fixed speed, they have the potential to resolve ambiguities in object scale and appearance. However, predicting depth end-to-end from sound and vision cannot benefit from unsupervised depth prediction approaches, which can process large scale data without sound annotation. In this work we show how Echoes can benefit depth prediction in two ways: When learning metric depth learned from supervised data and as supervisory signal for scale-correct self-supervised training. We show how we can improve the predictions of several state-of-the-art approaches and how the method can scale-correct a self-supervised depth approach.
Metric depth estimation from visual sensors is crucial for robots to perceive, navigate, and interact with their environment. Traditional range imaging setups, such as stereo or structured light cameras, face hassles including calibration, occlusions, and hardware demands, with accuracy limited by the baseline between cameras. Single- and multi-view monocular depth offers a more compact alternative, but is constrained by the unobservability of the metric scale. Light field imaging provides a promising solution for estimating metric depth by using a unique lens configuration through a single device. However, its application to single-view dense metric depth is under-addressed mainly due to the technology's high cost, the lack of public benchmarks, and proprietary geometrical models and software. Our work explores the potential of focused plenoptic cameras for dense metric depth. We propose a novel pipeline that predicts metric depth from a single plenoptic camera shot by first generating a sparse metric point cloud using machine learning, which is then used to scale and align a dense relative depth map regressed by a foundation depth model, resulting in dense metric depth. To validate it, we curated the Light Field & Stereo Image Dataset (LFS) of real-world light field images with stereo depth labels, filling a current gap in existing resources. Experimental results show that our pipeline produces accurate metric depth predictions, laying a solid groundwork for future research in this field.
Achieving robust stereo 3D imaging under diverse illumination conditions is an important however challenging task, due to the limited dynamic ranges (DRs) of cameras, which are significantly smaller than real world DR. As a result, the accuracy of existing stereo depth estimation methods is often compromised by under- or over-exposed images. Here, we introduce dual-exposure stereo for extended dynamic range 3D imaging. We develop automatic dual-exposure control method that adjusts the dual exposures, diverging them when the scene DR exceeds the camera DR, thereby providing information about broader DR. From the captured dual-exposure stereo images, we estimate depth using motion-aware dual-exposure stereo network. To validate our method, we develop a robot-vision system, collect stereo video datasets, and generate a synthetic dataset. Our method outperforms other exposure control methods.
Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions. Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a 69.5% improvement in accuracy over the previous SoTA on the ADIW dataset.
Tracking the 6DoF pose of unknown objects in monocular RGB video sequences is crucial for robotic manipulation. However, existing approaches typically rely on accurate depth information, which is non-trivial to obtain in real-world scenarios. Although depth estimation algorithms can be employed, geometric inaccuracy can lead to failures in RGBD-based pose tracking methods. To address this challenge, we introduce GSGTrack, a novel RGB-based pose tracking framework that jointly optimizes geometry and pose. Specifically, we adopt 3D Gaussian Splatting to create an optimizable 3D representation, which is learned simultaneously with a graph-based geometry optimization to capture the object's appearance features and refine its geometry. However, the joint optimization process is susceptible to perturbations from noisy pose and geometry data. Thus, we propose an object silhouette loss to address the issue of pixel-wise loss being overly sensitive to pose noise during tracking. To mitigate the geometric ambiguities caused by inaccurate depth information, we propose a geometry-consistent image pair selection strategy, which filters out low-confidence pairs and ensures robust geometric optimization. Extensive experiments on the OnePose and HO3D datasets demonstrate the effectiveness of GSGTrack in both 6DoF pose tracking and object reconstruction.
Generative models have significantly improved the generation and prediction quality on either camera images or LiDAR point clouds for autonomous driving. However, a real-world autonomous driving system uses multiple kinds of input modality, usually cameras and LiDARs, where they contain complementary information for generation, while existing generation methods ignore this crucial feature, resulting in the generated results only covering separate 2D or 3D information. In order to fill the gap in 2D-3D multi-modal joint generation for autonomous driving, in this paper, we propose our framework, \emph{HoloDrive}, to jointly generate the camera images and LiDAR point clouds. We employ BEV-to-Camera and Camera-to-BEV transform modules between heterogeneous generative models, and introduce a depth prediction branch in the 2D generative model to disambiguate the un-projecting from image space to BEV space, then extend the method to predict the future by adding temporal structure and carefully designed progressive training. Further, we conduct experiments on single frame generation and world model benchmarks, and demonstrate our method leads to significant performance gains over SOTA methods in terms of generation metrics.
Large Foundation Models like Dust3r can produce high quality outputs such as pointmaps, camera intrinsics, and depth estimation, given stereo-image pairs as input. However, the application of these outputs on tasks like Visual Localization requires a large amount of inference time and compute resources. To address these limitations, in this paper, we propose the use of a knowledge distillation pipeline, where we aim to build a student-teacher model with Dust3r as the teacher and explore multiple architectures of student models that are trained using the 3D reconstructed points output by Dust3r. Our goal is to build student models that can learn scene-specific representations and output 3D points with replicable performance such as Dust3r. The data set we used to train our models is 12Scenes. We test two main architectures of models: a CNN-based architecture and a Vision Transformer based architecture. For each architecture, we also compare the use of pre-trained models against models built from scratch. We qualitatively compare the reconstructed 3D points output by the student model against Dust3r's and discuss the various features learned by the student model. We also perform ablation studies on the models through hyperparameter tuning. Overall, we observe that the Vision Transformer presents the best performance visually and quantitatively.
In this work, we present SpaRC, a novel Sparse fusion transformer for 3D perception that integrates multi-view image semantics with Radar and Camera point features. The fusion of radar and camera modalities has emerged as an efficient perception paradigm for autonomous driving systems. While conventional approaches utilize dense Bird's Eye View (BEV)-based architectures for depth estimation, contemporary query-based transformers excel in camera-only detection through object-centric methodology. However, these query-based approaches exhibit limitations in false positive detections and localization precision due to implicit depth modeling. We address these challenges through three key contributions: (1) sparse frustum fusion (SFF) for cross-modal feature alignment, (2) range-adaptive radar aggregation (RAR) for precise object localization, and (3) local self-attention (LSA) for focused query aggregation. In contrast to existing methods requiring computationally intensive BEV-grid rendering, SpaRC operates directly on encoded point features, yielding substantial improvements in efficiency and accuracy. Empirical evaluations on the nuScenes and TruckScenes benchmarks demonstrate that SpaRC significantly outperforms existing dense BEV-based and sparse query-based detectors. Our method achieves state-of-the-art performance metrics of 67.1 NDS and 63.1 AMOTA. The code and pretrained models are available at https://github.com/phi-wol/sparc.
In underwater images, most useful features are occluded by water. The extent of the occlusion depends on imaging geometry and can vary even across a sequence of burst images. As a result, 3D reconstruction methods robust on in-air scenes, like Neural Radiance Field methods (NeRFs) or 3D Gaussian Splatting (3DGS), fail on underwater scenes. While a recent underwater adaptation of NeRFs achieved state-of-the-art results, it is impractically slow: reconstruction takes hours and its rendering rate, in frames per second (FPS), is less than 1. Here, we present a new method that takes only a few minutes for reconstruction and renders novel underwater scenes at 140 FPS. Named Gaussian Splashing, our method unifies the strengths and speed of 3DGS with an image formation model for capturing scattering, introducing innovations in the rendering and depth estimation procedures and in the 3DGS loss function. Despite the complexities of underwater adaptation, our method produces images at unparalleled speeds with superior details. Moreover, it reveals distant scene details with far greater clarity than other methods, dramatically improving reconstructed and rendered images. We demonstrate results on existing datasets and a new dataset we have collected. Additional visual results are available at: https://bgu-cs-vil.github.io/gaussiansplashingUW.github.io/ .