Reconstructing accurate 3D scenes from images is a long-standing vision task. Due to the ill-posedness of the single-image reconstruction problem, most well-established methods are built upon multi-view geometry. State-of-the-art (SOTA) monocular metric depth estimation methods can only handle a single camera model and are unable to perform mixed-data training due to the metric ambiguity. Meanwhile, SOTA monocular methods trained on large mixed datasets achieve zero-shot generalization by learning affine-invariant depths, which cannot recover real-world metrics. In this work, we show that the key to a zero-shot single-view metric depth model lies in the combination of large-scale data training and resolving the metric ambiguity from various camera models. We propose a canonical camera space transformation module, which explicitly addresses the ambiguity problems and can be effortlessly plugged into existing monocular models. Equipped with our module, monocular models can be stably trained with over 8 million images with thousands of camera models, resulting in zero-shot generalization to in-the-wild images with unseen camera settings. Experiments demonstrate SOTA performance of our method on 7 zero-shot benchmarks. Notably, our method won the championship in the 2nd Monocular Depth Estimation Challenge. Our method enables the accurate recovery of metric 3D structures on randomly collected internet images, paving the way for plausible single-image metrology. The potential benefits extend to downstream tasks, which can be significantly improved by simply plugging in our model. For example, our model relieves the scale drift issues of monocular-SLAM (Fig. 1), leading to high-quality metric scale dense mapping. The code is available at https://github.com/YvanYin/Metric3D.
Self-supervised monocular depth estimation (SS-MDE) has the potential to scale to vast quantities of data. Unfortunately, existing approaches limit themselves to the automotive domain, resulting in models incapable of generalizing to complex environments such as natural or indoor settings. To address this, we propose a large-scale SlowTV dataset curated from YouTube, containing an order of magnitude more data than existing automotive datasets. SlowTV contains 1.7M images from a rich diversity of environments, such as worldwide seasonal hiking, scenic driving and scuba diving. Using this dataset, we train an SS-MDE model that provides zero-shot generalization to a large collection of indoor/outdoor datasets. The resulting model outperforms all existing SSL approaches and closes the gap on supervised SoTA, despite using a more efficient architecture. We additionally introduce a collection of best-practices to further maximize performance and zero-shot generalization. This includes 1) aspect ratio augmentation, 2) camera intrinsic estimation, 3) support frame randomization and 4) flexible motion estimation. Code is available at https://github.com/jspenmar/slowtv_monodepth.
Learning-based multi-view stereo (MVS) methods deal with predicting accurate depth maps to achieve an accurate and complete 3D representation. Despite the excellent performance, existing methods ignore the fact that a suitable depth geometry is also critical in MVS. In this paper, we demonstrate that different depth geometries have significant performance gaps, even using the same depth prediction error. Therefore, we introduce an ideal depth geometry composed of Saddle-Shaped Cells, whose predicted depth map oscillates upward and downward around the ground-truth surface, rather than maintaining a continuous and smooth depth plane. To achieve it, we develop a coarse-to-fine framework called Dual-MVSNet (DMVSNet), which can produce an oscillating depth plane. Technically, we predict two depth values for each pixel (Dual-Depth), and propose a novel loss function and a checkerboard-shaped selecting strategy to constrain the predicted depth geometry. Compared to existing methods,DMVSNet achieves a high rank on the DTU benchmark and obtains the top performance on challenging scenes of Tanks and Temples, demonstrating its strong performance and generalization ability. Our method also points to a new research direction for considering depth geometry in MVS.
We introduce the notion of point affiliation into feature upsampling. By abstracting a feature map into non-overlapped semantic clusters formed by points of identical semantic meaning, feature upsampling can be viewed as point affiliation -- designating a semantic cluster for each upsampled point. In the framework of kernel-based dynamic upsampling, we show that an upsampled point can resort to its low-res decoder neighbors and high-res encoder point to reason the affiliation, conditioned on the mutual similarity between them. We therefore present a generic formulation for generating similarity-aware upsampling kernels and prove that such kernels encourage not only semantic smoothness but also boundary sharpness. This formulation constitutes a novel, lightweight, and universal upsampling solution, Similarity-Aware Point Affiliation (SAPA). We show its working mechanism via our preliminary designs with window-shape kernel. After probing the limitations of the designs on object detection, we reveal additional insights for upsampling, leading to SAPA with the dynamic kernel shape. Extensive experiments demonstrate that SAPA outperforms prior upsamplers and invites consistent performance improvements on a number of dense prediction tasks, including semantic segmentation, object detection, instance segmentation, panoptic segmentation, image matting, and depth estimation. Code is made available at: https://github.com/tiny-smart/sapa
Video depth estimation aims to infer temporally consistent depth. One approach is to finetune a single-image model on each video with geometry constraints, which proves inefficient and lacks robustness. An alternative is learning to enforce consistency from data, which requires well-designed models and sufficient video depth data. To address both challenges, we introduce NVDS+ that stabilizes inconsistent depth estimated by various single-image models in a plug-and-play manner. We also elaborate a large-scale Video Depth in the Wild (VDW) dataset, which contains 14,203 videos with over two million frames, making it the largest natural-scene video depth dataset. Additionally, a bidirectional inference strategy is designed to improve consistency by adaptively fusing forward and backward predictions. We instantiate a model family ranging from small to large scales for different applications. The method is evaluated on VDW dataset and three public benchmarks. To further prove the versatility, we extend NVDS+ to video semantic segmentation and several downstream applications like bokeh rendering, novel view synthesis, and 3D reconstruction. Experimental results show that our method achieves significant improvements in consistency, accuracy, and efficiency. Our work serves as a solid baseline and data foundation for learning-based video depth estimation. Code and dataset are available at: https://github.com/RaymondWang987/NVDS
Despite the increasing prevalence of rotating-style capture (e.g., surveillance cameras), conventional stereo rectification techniques frequently fail due to the rotation-dominant motion and small baseline between views. In this paper, we tackle the challenge of performing stereo rectification for uncalibrated rotating cameras. To that end, we propose Depth-from-Rotation (DfR), a novel image rectification solution that analytically rectifies two images with two-point correspondences and serves for further depth estimation. Specifically, we model the motion of a rotating camera as the camera rotates on a sphere with fixed latitude. The camera's optical axis lies perpendicular to the sphere's surface. We call this latitudinal motion assumption. Then we derive a 2-point analytical solver from directly computing the rectified transformations on the two images. We also present a self-adaptive strategy to reduce the geometric distortion after rectification. Extensive synthetic and real data experiments demonstrate that the proposed method outperforms existing works in effectiveness and efficiency by a significant margin.
Neural Radiance Field (NeRF) has been a mainstream in novel view synthesis with its remarkable quality of rendered images and simple architecture. Although NeRF has been developed in various directions improving continuously its performance, the necessity of a dense set of multi-view images still exists as a stumbling block to progress for practical application. In this work, we propose FlipNeRF, a novel regularization method for few-shot novel view synthesis by utilizing our proposed flipped reflection rays. The flipped reflection rays are explicitly derived from the input ray directions and estimated normal vectors, and play a role of effective additional training rays while enabling to estimate more accurate surface normals and learn the 3D geometry effectively. Since the surface normal and the scene depth are both derived from the estimated densities along a ray, the accurate surface normal leads to more exact depth estimation, which is a key factor for few-shot novel view synthesis. Furthermore, with our proposed Uncertainty-aware Emptiness Loss and Bottleneck Feature Consistency Loss, FlipNeRF is able to estimate more reliable outputs with reducing floating artifacts effectively across the different scene structures, and enhance the feature-level consistency between the pair of the rays cast toward the photo-consistent pixels without any additional feature extractor, respectively. Our FlipNeRF achieves the SOTA performance on the multiple benchmarks across all the scenarios.
Monocular depth estimation is scale-ambiguous, and thus requires scale supervision to produce metric predictions. Even so, the resulting models will be geometry-specific, with learned scales that cannot be directly transferred across domains. Because of that, recent works focus instead on relative depth, eschewing scale in favor of improved up-to-scale zero-shot transfer. In this work we introduce ZeroDepth, a novel monocular depth estimation framework capable of predicting metric scale for arbitrary test images from different domains and camera parameters. This is achieved by (i) the use of input-level geometric embeddings that enable the network to learn a scale prior over objects; and (ii) decoupling the encoder and decoder stages, via a variational latent representation that is conditioned on single frame information. We evaluated ZeroDepth targeting both outdoor (KITTI, DDAD, nuScenes) and indoor (NYUv2) benchmarks, and achieved a new state-of-the-art in both settings using the same pre-trained model, outperforming methods that train on in-domain data and require test-time scaling to produce metric estimates.
Dense pixel-specific representation learning at scale has been bottlenecked due to the unavailability of large-scale multi-view datasets. Current methods for building effective pretraining datasets heavily rely on annotated 3D meshes, point clouds, and camera parameters from simulated environments, preventing them from building datasets from real-world data sources where such metadata is lacking. We propose a pretraining dataset-curation approach that does not require any additional annotations. Our method allows us to generate multi-view datasets from both real-world videos and simulated environments at scale. Specifically, we experiment with two scales: MIMIC-1M with 1.3M and MIMIC-3M with 3.1M multi-view image pairs. We train multiple models with different masked image modeling objectives to showcase the following findings: Representations trained on our automatically generated MIMIC-3M outperform those learned from expensive crowdsourced datasets (ImageNet-1K) and those learned from synthetic environments (MULTIVIEW-HABITAT) on two dense geometric tasks: depth estimation on NYUv2 (1.7%), and surface normals estimation on Taskonomy (2.05%). For dense tasks which also require object understanding, we outperform MULTIVIEW-HABITAT, on semantic segmentation on ADE20K (3.89%), pose estimation on MSCOCO (9.4%), and reduce the gap with models pre-trained on the object-centric expensive ImageNet-1K. We outperform even when the representations are frozen, and when downstream training data is limited to few-shot. Larger dataset (MIMIC-3M) significantly improves performance, which is promising since our curation method can arbitrarily scale to produce even larger datasets. MIMIC code, dataset, and pretrained models are open-sourced at https://github.com/RAIVNLab/MIMIC.
Depth completion is the task of recovering dense depth maps from sparse ones, usually with the help of color images. Existing image-guided methods perform well on daytime depth perception self-driving benchmarks, but struggle in nighttime scenarios with poor visibility and complex illumination. To address these challenges, we propose a simple yet effective framework called LDCNet. Our key idea is to use Recurrent Inter-Convolution Differencing (RICD) and Illumination-Affinitive Intra-Convolution Differencing (IAICD) to enhance the nighttime color images and reduce the negative effects of the varying illumination, respectively. RICD explicitly estimates global illumination by differencing two convolutions with different kernels, treating the small-kernel-convolution feature as the center of the large-kernel-convolution feature in a new perspective. IAICD softly alleviates local relative light intensity by differencing a single convolution, where the center is dynamically aggregated based on neighboring pixels and the estimated illumination map in RICD. On both nighttime depth completion and depth estimation tasks, extensive experiments demonstrate the effectiveness of our LDCNet, reaching the state of the art.
Defocus blur detection (DBD) separates in-focus and out-of-focus regions in an image. Previous approaches mistakenly mistook homogeneous areas in focus for defocus blur regions, likely due to not considering the internal factors that cause defocus blur. Inspired by the law of depth, depth of field (DOF), and defocus, we propose an approach called D-DFFNet, which incorporates depth and DOF cues in an implicit manner. This allows the model to understand the defocus phenomenon in a more natural way. Our method proposes a depth feature distillation strategy to obtain depth knowledge from a pre-trained monocular depth estimation model and uses a DOF-edge loss to understand the relationship between DOF and depth. Our approach outperforms state-of-the-art methods on public benchmarks and a newly collected large benchmark dataset, EBD. Source codes and EBD dataset are available at: https:github.com/yuxinjin-whu/D-DFFNet.
Recent camera-based 3D object detection methods have introduced sequential frames to improve the detection performance hoping that multiple frames would mitigate the large depth estimation error. Despite improved detection performance, prior works rely on naive fusion methods (e.g., concatenation) or are limited to static scenes (e.g., temporal stereo), neglecting the importance of the motion cue of objects. These approaches do not fully exploit the potential of sequential images and show limited performance improvements. To address this limitation, we propose a novel 3D object detection model, P2D (Predict to Detect), that integrates a prediction scheme into a detection framework to explicitly extract and leverage motion features. P2D predicts object information in the current frame using solely past frames to learn temporal motion features. We then introduce a novel temporal feature aggregation method that attentively exploits Bird's-Eye-View (BEV) features based on predicted object information, resulting in accurate 3D object detection. Experimental results demonstrate that P2D improves mAP and NDS by 3.0% and 3.7% compared to the sequential image-based baseline, illustrating that incorporating a prediction scheme can significantly improve detection accuracy.
Exploring robust and efficient association methods has always been an important issue in multiple-object tracking (MOT). Although existing tracking methods have achieved impressive performance, congestion and frequent occlusions still pose challenging problems in multi-object tracking. We reveal that performing sparse decomposition on dense scenes is a crucial step to enhance the performance of associating occluded targets. To this end, we propose a pseudo-depth estimation method for obtaining the relative depth of targets from 2D images. Secondly, we design a depth cascading matching (DCM) algorithm, which can use the obtained depth information to convert a dense target set into multiple sparse target subsets and perform data association on these sparse target subsets in order from near to far. By integrating the pseudo-depth method and the DCM strategy into the data association process, we propose a new tracker, called SparseTrack. SparseTrack provides a new perspective for solving the challenging crowded scene MOT problem. Only using IoU matching, SparseTrack achieves comparable performance with the state-of-the-art (SOTA) methods on the MOT17 and MOT20 benchmarks. Code and models are publicly available at \url{https://github.com/hustvl/SparseTrack}.
When used by autonomous vehicles for trajectory planning or obstacle avoidance, depth estimation methods need to be reliable. Therefore, estimating the quality of the depth outputs is critical. In this paper, we show how M4Depth, a state-of-the-art depth estimation method designed for unmanned aerial vehicle (UAV) applications, can be enhanced to perform joint depth and uncertainty estimation. For that, we present a solution to convert the uncertainty estimates related to parallax generated by M4Depth into uncertainty estimates related to depth, and show that it outperforms the standard probabilistic approach. Our experiments on various public datasets demonstrate that our method performs consistently, even in zero-shot transfer. Besides, our method offers a compelling value when compared to existing multi-view depth estimation methods as it performs similarly on a multi-view depth estimation benchmark despite being 2.5 times faster and causal, as opposed to other methods. The code of our method is publicly available at https://github.com/michael-fonder/M4DepthU .
Decoders play significant roles in recovering scene depths. However, the decoders used in previous works ignore the propagation of multilevel lossless fine-grained information, cannot adaptively capture local and global information in parallel, and cannot perform sufficient global statistical analyses on the final output disparities. In addition, the process of mapping from a low-resolution feature space to a high-resolution feature space is a one-to-many problem that may have multiple solutions. Therefore, the quality of the recovered depth map is low. To this end, we propose a high-quality decoder (HQDec), with which multilevel near-lossless fine-grained information, obtained by the proposed adaptive axial-normalized position-embedded channel attention sampling module (AdaAxialNPCAS), can be adaptively incorporated into a low-resolution feature map with high-level semantics utilizing the proposed adaptive information exchange scheme. In the HQDec, we leverage the proposed adaptive refinement module (AdaRM) to model the local and global dependencies between pixels in parallel and utilize the proposed disparity attention module to model the distribution characteristics of disparity values from a global perspective. To recover fine-grained high-resolution features with maximal accuracy, we adaptively fuse the high-frequency information obtained by constraining the upsampled solution space utilizing the local and global dependencies between pixels into the high-resolution feature map generated from the nonlearning method. Extensive experiments demonstrate that each proposed component improves the quality of the depth estimation results over the baseline results, and the developed approach achieves state-of-the-art results on the KITTI and DDAD datasets. The code and models will be publicly available at \href{https://github.com/fwucas/HQDec}{HQDec}.
Neural radiance fields (NeRF) shows powerful performance in novel view synthesis and 3D geometry reconstruction, but it suffers from critical performance degradation when the number of known viewpoints is drastically reduced. Existing works attempt to overcome this problem by employing external priors, but their success is limited to certain types of scenes or datasets. Employing monocular depth estimation (MDE) networks, pretrained on large-scale RGB-D datasets, with powerful generalization capability would be a key to solving this problem: however, using MDE in conjunction with NeRF comes with a new set of challenges due to various ambiguity problems exhibited by monocular depths. In this light, we propose a novel framework, dubbed D\"aRF, that achieves robust NeRF reconstruction with a handful of real-world images by combining the strengths of NeRF and monocular depth estimation through online complementary training. Our framework imposes the MDE network's powerful geometry prior to NeRF representation at both seen and unseen viewpoints to enhance its robustness and coherence. In addition, we overcome the ambiguity problems of monocular depths through patch-wise scale-shift fitting and geometry distillation, which adapts the MDE network to produce depths aligned accurately with NeRF geometry. Experiments show our framework achieves state-of-the-art results both quantitatively and qualitatively, demonstrating consistent and reliable performance in both indoor and outdoor real-world datasets. Project page is available at https://ku-cvlab.github.io/DaRF/.
Light field (LF) depth estimation is a crucial task with numerous practical applications. However, mainstream methods based on the multi-view stereo (MVS) are resource-intensive and time-consuming as they need to construct a finer cost volume. To address this issue and achieve a better trade-off between accuracy and efficiency, we propose an occlusion-aware cascade cost volume for LF depth (disparity) estimation. Our cascaded strategy reduces the sampling number while keeping the sampling interval constant during the construction of a finer cost volume. We also introduce occlusion maps to enhance accuracy in constructing the occlusion-aware cost volume. Specifically, we first obtain the coarse disparity map through the coarse disparity estimation network. Then, the sub-aperture images (SAIs) of side views are warped to the center view based on the initial disparity map. Next, we propose photo-consistency constraints between the warped SAIs and the center SAI to generate occlusion maps for each SAI. Finally, we introduce the coarse disparity map and occlusion maps to construct an occlusion-aware refined cost volume, enabling the refined disparity estimation network to yield a more precise disparity map. Extensive experiments demonstrate the effectiveness of our method. Compared with state-of-the-art methods, our method achieves a superior balance between accuracy and efficiency and ranks first in terms of MSE and Q25 metrics among published methods on the HCI 4D benchmark. The code and model of the proposed method are available at https://github.com/chaowentao/OccCasNet.
We propose SING (StabIlized and Normalized Gradient), a plug-and-play technique that improves the stability and generalization of the Adam(W) optimizer. SING is straightforward to implement and has minimal computational overhead, requiring only a layer-wise standardization of the gradients fed to Adam(W) without introducing additional hyper-parameters. We support the effectiveness and practicality of the proposed approach by showing improved results on a wide range of architectures, problems (such as image classification, depth estimation, and natural language processing), and in combination with other optimizers. We provide a theoretical analysis of the convergence of the method, and we show that by virtue of the standardization, SING can escape local minima narrower than a threshold that is inversely proportional to the network's depth.
Text-driven 3D scene generation is widely applicable to video gaming, film industry, and metaverse applications that have a large demand for 3D scenes. However, existing text-to-3D generation methods are limited to producing 3D objects with simple geometries and dreamlike styles that lack realism. In this work, we present Text2NeRF, which is able to generate a wide range of 3D scenes with complicated geometric structures and high-fidelity textures purely from a text prompt. To this end, we adopt NeRF as the 3D representation and leverage a pre-trained text-to-image diffusion model to constrain the 3D reconstruction of the NeRF to reflect the scene description. Specifically, we employ the diffusion model to infer the text-related image as the content prior and use a monocular depth estimation method to offer the geometric prior. Both content and geometric priors are utilized to update the NeRF model. To guarantee textured and geometric consistency between different views, we introduce a progressive scene inpainting and updating strategy for novel view synthesis of the scene. Our method requires no additional training data but only a natural language description of the scene as the input. Extensive experiments demonstrate that our Text2NeRF outperforms existing methods in producing photo-realistic, multi-view consistent, and diverse 3D scenes from a variety of natural language prompts. Our code is available at https://github.com/eckertzhang/Text2NeRF.
For augmented reality (AR), it is important that virtual assets appear to `sit among' real world objects. The virtual element should variously occlude and be occluded by real matter, based on a plausible depth ordering. This occlusion should be consistent over time as the viewer's camera moves. Unfortunately, small mistakes in the estimated scene depth can ruin the downstream occlusion mask, and thereby the AR illusion. Especially in real-time settings, depths inferred near boundaries or across time can be inconsistent. In this paper, we challenge the need for depth-regression as an intermediate step. We instead propose an implicit model for depth and use that to predict the occlusion mask directly. The inputs to our network are one or more color images, plus the known depths of any virtual geometry. We show how our occlusion predictions are more accurate and more temporally stable than predictions derived from traditional depth-estimation models. We obtain state-of-the-art occlusion results on the challenging ScanNetv2 dataset and superior qualitative results on real scenes.