Dense Depth Map Research Articles

Real-time, dense UAV mapping by leveraging monocular depth prediction with monocular-inertial SLAM

We present a dense and metric 3D mapping pipeline designed for embedded operation on-board UAVs, by loosely coupling deep neural networks trained to infer dense depth single images with a SLAM system that restores metric scale from sparse depth. In contrast to computationally restrictive approaches that leverage multiple views, we propose a highly efficient, single-view approach without sacrificing 3D mapping performance. This enables real-time construction of a global 3D voxel map by iterative fusion of the rescaled dense depth maps obtained via raycasting from the estimated camera poses. Quantitative and qualitative experimentations of our framework in challenging environmental conditions show comparable or superior performance with respect to state-of-the-art approaches via a better effectiveness-efficiency trade-off.

Dense projection fusion for 3D object detection

Fusing information from LiDAR and cameras can effectively enhance the overall perceptivity of autonomous vehicles in various scenarios. Despite the relatively good results achieved by point-wise fusion and Bird’s-Eye-View (BEV) fusion, they still cannot fully leverage the image information and lack of effective depth information. For any fusion methods, the multi-modal features first need to be concatenated along the channel, and then the fused features are extracted using convolutional layers. However, this type of fusion methods is effective, but too coarse which causes that the fused features cannot pay more attention to the regions with important features and suffer from severe noise. To tackle these issues, we propose in this paper a Dense Projection Fusion (DPFusion) approach. It consists of two new modules: dense depth map guided BEV transform (DGBT) module and multi-modal feature adaptive fusion (MFAF) module. The DGBT module first quickly estimates the depth of each pixel and then projects all image features to the BEV space, making full use of the image information. The MFAF module computes the image weights and point cloud weights for each channel in each BEV grid and then adaptively weights and fuses the image BEV features with the point cloud BEV features. It is worth pointing out that the MFAF module makes the fused features pay more attention to background outline and object outline. Our proposed DPFusion demonstrates competitive results in 3D object detection, achieving a mean Average Precision (mAP) of 70.4 and a nuScenes detection score (NDS) of 72.3 on the nuScenes validation set.

Reconstruction of three-dimensional scenes based on video flow data

This work is dedicated to the application of modern algorithms for reconstructing spatial scenes from images to restore spatial information from video. The work is looking at a variety of modern methods, approaches, algorithms and trends in the field. The attention was paid to the sequence of development of approaches to the completion of the task. While researching the field and results related to three-dimensional reconstruction based on images and video streams, an algorithm was invented that allows constructing dense depth maps using information from all video frames. The idea is to use ready-made, commonly accepted, and tested solutions to solve two problems: COLMAP for visual odometry, and RAFT for computing optical flow. The algorithm shows quite accurate results and reconstructs the depth map in detail on arbitrary static scenes.

CAFNet: Context aligned fusion for depth completion

Depth completion aims at reconstructing a dense depth from sparse depth input, frequently using color images as guidance. The sparse depth map lacks sufficient contexts for reconstructing focal contexts such as the shape of objects. The RGB images contain redundant contexts including details useless for reconstruction, which reduces the efficiency of focal context extraction. The unaligned contextual information from these two modalities poses a challenge to focal context extraction and further fusion, as well as the accuracy of depth completion. To optimize the utilization of multimodal contextual information, we explore a novel framework: Context Aligned Fusion Network (CAFNet). CAFNet comprises two stages: the context-aligned stage and the full-scale stage. In the context-aligned stage, CAFNet downsamples input RGB-D pairs to the scale, at which multimodal contextual information is adequately aligned for feature extraction in two encoders and fusion in CF modules. In the full-scale stage, feature maps with fused multimodal context from the previous stage are upsampled to the original scale and subsequentially fused with full-scale depth features by the GF module utilizing a dynamic masked fusion strategy. Ultimately, accurate dense depth maps are reconstructed, leveraging the GF module’s resultant features. Experiments conducted on indoor and outdoor benchmark datasets show that the CAFNet produces results comparable to state-of-the-art methods while effectively reducing computational costs.

Learnable differencing center for nighttime depth perception

Depth completion is the task of recovering dense depth map 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 learnable differencing center network (LDCNet). The 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 the 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 the RICD. On both nighttime depth completion and depth estimation tasks, extensive experiments demonstrate the effectiveness of our LDCNet, reaching the state of the art.

Learning Cross-modality Interaction for Robust Depth Perception of Autonomous Driving

As one of the fundamental tasks of autonomous driving, depth perception aims to perceive physical objects in three dimensions and to judge their distances away from the ego vehicle. Although great efforts have been made for depth perception, LiDAR-based and camera-based solutions have limitations with low accuracy and poor robustness for noise input. With the integration of monocular cameras and LiDAR sensors in autonomous vehicles, in this article, we introduce a two-stream architecture to learn the modality interaction representation under the guidance of an image reconstruction task to compensate for the deficiencies of each modality in a parallel manner. Specifically, in the two-stream architecture, the multi-scale cross-modality interactions are preserved via a cascading interaction network under the guidance of the reconstruction task. Next, the shared representation of modality interaction is integrated to infer the dense depth map due to the complementarity and heterogeneity of the two modalities. We evaluated the proposed solution on the KITTI dataset and CALAR synthetic dataset. Our experimental results show that learning the coupled interaction of modalities under the guidance of an auxiliary task can lead to significant performance improvements. Furthermore, our approach is competitive against the state-of-the-art models and robust against the noisy input. The source code is available at https://github.com/tonyFengye/Code/tree/master .

Recent Advances in Conventional and Deep Learning-Based Depth Completion: A Survey.

Depth completion aims to recover pixelwise depth from incomplete and noisy depth measurements with or without the guidance of a reference RGB image. This task attracted considerable research interest due to its importance in various computer vision-based applications, such as scene understanding, autonomous driving, 3-D reconstruction, object detection, pose estimation, trajectory prediction, and so on. As the system input, an incomplete depth map is usually generated by projecting the 3-D points collected by ranging sensors, such as LiDAR in outdoor environments, or obtained directly from RGB-D cameras in indoor areas. However, even if a high-end LiDAR is employed, the obtained depth maps are still very sparse and noisy, especially in the regions near the object boundaries, which makes the depth completion task a challenging problem. To address this issue, a few years ago, conventional image processing-based techniques were employed to fill the holes and remove the noise from the relatively dense depth maps obtained by RGB-D cameras, while deep learning-based methods have recently become increasingly popular and inspiring results have been achieved, especially for the challenging situation of LiDAR-image-based depth completion. This article systematically reviews and summarizes the works related to the topic of depth completion in terms of input modalities, data fusion strategies, loss functions, and experimental settings, especially for the key techniques proposed in deep learning-based multiple input methods. On this basis, we conclude by presenting the current status of depth completion and discussing several prospects for its future research directions.

METER: A Mobile Vision Transformer Architecture for Monocular Depth Estimation

Depth estimation is a fundamental knowledge for autonomous systems that need to assess their own state and perceive the surrounding environment. Deep learning algorithms for depth estimation have gained significant interest in recent years, owing to the potential benefits of this methodology in overcoming the limitations of active depth sensing systems. Moreover, due to the low cost and size of monocular cameras, researchers have focused their attention on monocular depth estimation (MDE), which consists in estimating a dense depth map from a single RGB video frame. State of the art MDE models typically rely on vision transformers (ViT) architectures that are highly deep and complex, making them unsuitable for fast inference on devices with hardware constraints. Purposely, in this paper, we address the problem of exploiting ViT in MDE on embedded devices. Those systems are usually characterized by limited memory capabilities and low-power CPU/GPU. We propose METER, a novel lightweight vision transformer architecture capable of achieving state of the art estimations and low latency inference performances on the considered embedded hardwares: NVIDIA Jetson TX1 and NVIDIA Jetson Nano. We provide a solution consisting of three alternative configurations of METER, a novel loss function to balance pixel estimation and reconstruction of image details, and a new data augmentation strategy to improve the overall final predictions. The proposed method outperforms previous lightweight works over the two benchmark datasets: the indoor NYU Depth v2 and the outdoor KITTI.

Radar-Camera Fusion Network for Depth Estimation in Structured Driving Scenes.

Depth estimation is an important part of the perception system in autonomous driving. Current studies often reconstruct dense depth maps from RGB images and sparse depth maps obtained from other sensors. However, existing methods often pay insufficient attention to latent semantic information. Considering the highly structured characteristics of driving scenes, we propose a dual-branch network to predict dense depth maps by fusing radar and RGB images. The driving scene is divided into three parts in the proposed architecture, each predicting a depth map, which is finally merged into one by implementing the fusion strategy in order to make full use of the potential semantic information in the driving scene. In addition, a variant L1 loss function is applied in the training phase, directing the network to focus more on those areas of interest when driving. Our proposed method is evaluated on the nuScenes dataset. Experiments demonstrate its effectiveness in comparison with previous state of the art methods.

Micro 4D Imaging Sensor Using Snapshot Narrowband Imaging Method.

The spectral and depth (SAD) imaging method plays an important role in the field of computer vision. However, accurate depth estimation and spectral image capture from a single image without increasing the volume of the imaging sensor is still an unresolved problem. Our research finds that a snapshot narrow band imaging (SNBI) method can discern wavelength-dependent spectral aberration and simultaneously capture spectral-aberration defocused images for quantitative depth estimation. First, a micro 4D imaging (M4DI) sensor is proposed by integrating a mono-chromatic imaging sensor with a miniaturized narrow-band microarrayed spectral filter mosaic. The appearance and volume of the M4DI sensor are the same as the integrated mono-chromatic imaging sensor. A simple remapping algorithm was developed to separate the raw image into four narrow spectral band images. Then, a depth estimation algorithm is developed to generate 3D data with a dense depth map at every exposure of the M4DI sensor. Compared with existing SAD imaging method, the M4DI sensor has the advantages of simple implementation, low computational burden, and low cost. A proof-of-principle M4DI sensor was applied to sense the depth of objects and to track a tiny targets trajectory. The relative error in the three-dimensional positioning is less than 7% for objects within 1.1 to 2.8 m.

Non-local affinity adaptive acceleration propagation network for generating dense depth maps from LiDAR.

Depth completion aims to generate dense depth maps from the sparse depth images generated by LiDAR. In this paper, we propose a non-local affinity adaptive accelerated (NL-3A) propagation network for depth completion to solve the mixing depth problem of different objects on the depth boundary. In the network, we design the NL-3A prediction layer to predict the initial dense depth maps and their reliability, non-local neighbors and affinities of each pixel, and learnable normalization factors. Compared with the traditional fixed-neighbor affinity refinement scheme, the non-local neighbors predicted by the network can overcome the propagation error problem of mixed depth objects. Subsequently, we combine the learnable normalized propagation of non-local neighbor affinity with pixel depth reliability in the NL-3A propagation layer, so that it can adaptively adjust the propagation weight of each neighbor during the propagation process, which enhances the robustness of the network. Finally, we design an accelerated propagation model. This model enables parallel propagation of all neighbor affinities and improves the efficiency of refining dense depth maps. Experiments on KITTI depth completion and NYU Depth V2 datasets show that our network is superior to most algorithms in terms of accuracy and efficiency of depth completion. In particular, we predict and reconstruct more smoothly and consistently at the pixel edges of different objects.

H-Seg: a horizontal reconstruction volume segmentation method for accurate depth estimation in a computer-generated hologram.

In this work, we introduce a novel approach for depth estimation in a computer-generated hologram by employing horizontal segmentation of the reconstruction volume instead of conventional vertical segmentation. The reconstruction volume is divided into horizontal slices and each slice is processed using a residual U-net architecture to identify in-focus lines, enabling determination of the slice's intersection with the 3D scene. The individual slice results are then combined to generate a dense depth map of the scene. Our experiments demonstrate the effectiveness of our method, with improved accuracy, faster processing times, lower graphics processing unit (GPU) utilization, and smoother predicted depth maps than existing state-of-the-art models.

3-D Scene Flow Estimation on Pseudo-LiDAR: Bridging the Gap on Estimating Point Motion

3D scene flow characterizes how the points at the current time flow to the next time in the 3D Euclidean space, which possesses the capacity to infer autonomously the non-rigid motion of all objects in the scene. The previous methods for estimating scene flow from images have limitations, which split the holistic nature of 3D scene flow by estimating optical flow and disparity separately. Learning 3D scene flow from point clouds also faces the difficulties of the gap between synthesized and real data and the sparsity of LiDAR point clouds. In this paper, the generated dense depth map is utilized to obtain explicit 3D coordinates, which achieves direct learning of 3D scene flow from 2D images. The stability of the predicted scene flow is improved by introducing the dense nature of 2D pixels into the 3D space. Outliers in the generated 3D point cloud are removed by statistical methods to weaken the impact of noisy points on the 3D scene flow estimation task. Disparity consistency loss is proposed to achieve more effective unsupervised learning of 3D scene flow. The proposed method of self-supervised learning of 3D scene flow on real-world images is compared with a variety of methods for learning on the synthesized dataset and learning on LiDAR point clouds. The comparisons of multiple scene flow metrics are shown to demonstrate the effectiveness and superiority of introducing pseudo-LiDAR point cloud to scene flow estimation.

Self‐supervised depth completion with multi‐view geometric constraints

AbstractSelf‐supervised learning‐based depth completion is a cost‐effective way for 3D environment perception. However, it is also a challenging task because sparse depth may deactivate neural networks. In this paper, a novel Sparse‐Dense Depth Consistency Loss (SDDCL) is proposed to penalize not only the estimated depth map with sparse input points but also consecutive completed dense depth maps. Combined with the pose consistency loss, a new self‐supervised learning scheme is developed, using multi‐view geometric constraints, to achieve more accurate depth completion results. Moreover, to tackle the sparsity issue of input depth, a Quasi Dense Representations (QDR) module with triplet branches for spatial pyramid pooling is proposed to produce more dense feature maps. Extensive experimental results on VOID, NYUv2, and KITTI datasets show that the method outperforms state‐of‐the‐art self‐supervised depth completion methods.

Sparse-to-dense coarse-to-fine depth estimation for colonoscopy

Colonoscopy, as the golden standard for screening colon cancer and diseases, offers considerable benefits to patients. However, it also imposes challenges on diagnosis and potential surgery due to the narrow observation perspective and limited perception dimension. Dense depth estimation can overcome the above limitations and offer doctors straightforward 3D visual feedback. To this end, we propose a novel sparse-to-dense coarse-to-fine depth estimation solution for colonoscopic scenes based on the direct SLAM algorithm. The highlight of our solution is that we utilize the scattered 3D points obtained from SLAM to generate accurate and dense depth in full resolution. This is done by a deep learning (DL)-based depth completion network and a reconstruction system. The depth completion network effectively extracts texture, geometry, and structure features from sparse depth along with RGB data to recover the dense depth map. The reconstruction system further updates the dense depth map using a photometric error-based optimization and a mesh modeling approach to reconstruct a more accurate 3D model of colons with detailed surface texture. We show the effectiveness and accuracy of our depth estimation method on near photo-realistic challenging colon datasets. Experiments demonstrate that the strategy of sparse-to-dense coarse-to-fine can significantly improve the performance of depth estimation and smoothly fuse direct SLAM and DL-based depth estimation into a complete dense reconstruction system.

Single-shot dense depth sensing with a tricolor RGB fringe pattern

A single-shot fringe pattern composed of red (R), green (G), and blue (B) is proposed for dynamic scene depth acquisition. A tricolor binary fringe is adopted and filtered in the frequency domain to obtain sinusoidal fringes. A Gabor filter is employed to estimate the wrapped phase. To overcome the phase ambiguity, we use the color-encoded De Bruijn sequence to denote the period of the phase information and assign the sequence to two channels of the pattern, while the third channel is equal to the inverse of the composition of the two channels. Benefiting from this coding strategy, two disparity maps with different phase shifts can be acquired in only one shot, and a disparity consistency check is used to eliminate errors. Quantitative and qualitative experiments are conducted to verify the performance of the proposed method and prove that it can achieve a high accuracy and dense depth map.

SPNet: Structure preserving network for depth completion.

Depth completion aims to predict a dense depth map from a sparse one. Benefiting from the powerful ability of convolutional neural networks, recent depth completion methods have achieved remarkable performance. However, it is still a challenging problem to well preserve accurate depth structures, such as tiny structures and object boundaries. To tackle this problem, we propose a structure preserving network (SPNet) in this paper. Firstly, an efficient multi-scale gradient extractor (MSGE) is proposed to extract useful multi-scale gradient images, which contain rich structural information that is helpful in recovering accurate depth. The MSGE is constructed based on the proposed semi-fixed depthwise separable convolution. Meanwhile, we adopt a stable gradient MAE loss (LGMAE) to provide additional depth gradient constrain for better structure reconstruction. Moreover, a multi-level feature fusion module (MFFM) is proposed to adaptively fuse the spatial details from low-level encoder and the semantic information from high-level decoder, which will incorporate more structural details into the depth modality. As demonstrated by experiments on NYUv2 and KITTI datasets, our method outperforms some state-of-the-art methods in terms of both quantitative and quantitative evaluations.

Spacecraft depth completion based on the gray image and the sparse depth map

Perceiving the three-dimensional (3D) structure of the spacecraft is a prerequisite for successfully executing many on-orbit space missions, and it can provide critical input for many downstream vision algorithms. In this paper, we propose to sense the 3D structure of spacecraft using light detection and ranging sensor (LIDAR) and a monocular camera. To this end, Spacecraft Depth Completion Network (SDCNet) is proposed to recover the dense depth map based on gray image and sparse depth map. Specifically, SDCNet decomposes the object-level spacecraft depth completion task into foreground segmentation subtask and foreground depth completion subtask, which segments the spacecraft region first and then performs depth completion on the segmented foreground area. In this way, the background interference to foreground spacecraft depth completion is effectively avoided. Moreover, an attention-based feature fusion module is also proposed to aggregate the complementary information between different inputs, which deduces the correlation between different features along the channel and the spatial dimension sequentially. Besides, four metrics are also proposed to evaluate object-level depth completion performance, which can more intuitively reflect the quality of spacecraft depth completion results. Finally, a large-scale satellite depth completion dataset is constructed for training and testing spacecraft depth completion algorithms. Empirical experiments on the dataset demonstrate the effectiveness of the proposed SDCNet, which achieves 0.225m mean absolute error of interest and 0.778m mean absolute truncation error, surpassing state-of-the-art methods by a large margin. The spacecraft pose estimation experiment is also conducted based on the depth completion results, and the experimental results indicate that the predicted dense depth map could meet the needs of downstream vision tasks.

Dense Depth-Guided Generalizable NeRF

Neural rendering approaches enable photo-realistic rendering on novel view synthesis tasks while their per-scene optimization remains an issue for scalability. Recent methods introduce novel neural radiance field (NeRF) frameworks that generalize to unseen scenes on-the-fly by combining multi-view stereo with differentiable volume rendering. These generalizable NeRF methods synthesize the colors of 3D ray points by learning the consistency of image features projected from given nearby views. Since the consistency is computed on the 2D projected image space, it is vulnerable to occlusion and local shape variation by viewing direction. To solve this problem, we present dense depth-guided generalizable NeRF that leverages the depth as the signed distance between the ray point and the object surface of the scene. We first generate the dense depth maps from sparse 3D points of structure from motion (SfM) which is an inevitable step to obtain camera poses. Next, the dense depth maps are exploited as complementary features invariant to the sparsity of nearby views and mask for occlusion handling. Experiments demonstrate that our approach outperforms existing generalizable NeRF methods for widely used real and synthetic datasets.

Nano Quadcopter Obstacle Avoidance with a Lightweight Monocular Depth Network

In this paper, we propose an obstacle avoidance solution for a 34-gram quadcopter equipped with a monocular camera. The perception of obstacles is tackled by a lightweight convolutional neural network predicting a dense depth map from a captured grey-scale image. The depth network performs self-supervised learning and thus requires no ground-truth labels that are costly to acquire. Based on the depth map, the control strategy is implemented by a behavior state machine that balances the efficiency to explore the environment and the safety of avoiding obstacles. In real-world flight experiments, our solution demonstrates the efficacy of predicting trust-worthy depth maps and a stable control strategy in various cluttered environments.