DASNeRF: depth consistency optimization, adaptive sampling, and hierarchical structural fusion for sparse view neural radiance fields

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(5), P. e0321878 - e0321878

Published: May 12, 2025

To address the challenges of significant detail loss in Neural Radiance Fields (NeRF) under sparse-view input conditions, this paper proposes DASNeRF framework. aims to generate high-detail novel views from a limited number viewpoints. limitations few-shot NeRF, including insufficient depth information and loss, introduces accurate priors employs constraint strategy combining relative ordering fidelity regularization structural consistency regularization. These methods ensure reconstruction accuracy even with sparse views. The provide high-quality data through more monocular estimation model, enhancing capability stability model. guides network learn relationships using local ranking priors, reducing blurring caused by inaccurate estimation. Depth maintains global enforcing continuity across neighboring pixels. strategies enhance DASNeRF’s performance complex scenes, making 3D natural. In addition, we utilize three-layer optimal sampling strategy, consisting coarse sampling, optimized fine during process better capture details key regions. phase, point density regions is adaptively increased while low-priority regions, accuracy. alleviate overfitting, proposed an MLP structure per-layer fusion. This design preserves model’s perception ability effectively avoids overfitting. Specifically, each layer’s includes output features previous layer incorporates processed five-dimensional information, further reconstruction. Experimental results show that outperforms state-of-the-art on LLFF DTU dataset, achieving metrics such as PSNR, SSIM, LPIPS. reconstructed visual quality are significantly improved, demonstrating potential conditions.

Language: Английский

---

Rapid depth estimation based on a key electrowetting liquid-lens with electrically adjusted imaging focus

Review of Scientific Instruments, Journal Year: 2025, Volume and Issue: 96(5)

Published: May 1, 2025

An effective method for rapidly performing depth estimation using a type of electrowetting liquid-lens is proposed. The architectured by directly coupling cylindrical copper sidewall electrode and top ITO electrode, leading to dual-mode lens adjusted electrically, including beam diverging mode converging mode, also an intermedium phase retard state. By increasing the applied signal voltage from 0 120 V, focus presents wide dynamic range (-∞, −128.6 mm) ∪ (45.6 mm, +∞). key performances liquid-lens, such as tunable electrically element response duration less than 5 ms, are evaluated experimentally. sweeping over coupled with arrayed CMOS sensor form imaging setup, sequence images focal stack acquired. Considering character field equipment mainly based on can be remarkably extended further utilizing effect in transition region between positive negative focus. A rapid algorithm aligning then eliminating scene parallax achieved.

Language: Английский

---

Fully Self-Supervised Depth Estimation from Defocus Clue

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2023, Volume and Issue: unknown, P. 9140 - 9149

Published: June 1, 2023

Depth-from-defocus (DFD), modeling the relationship between depth and defocus pattern in images, has demonstrated promising performance estimation. Recently, several self-supervised works try to overcome difficulties acquiring accurate ground-truth. However, they depend on all-in-focus (AIF) which cannot be captured real-world scenarios. Such limitation discourages applications of DFD methods. To tackle this issue, we propose a completely framework that estimates purely from sparse focal stack. We show our circumvents needs for AIF image ground-truth, receives superior predictions, thus closing gap theoretical success their real world. In particular, (i) more realistic setting tasks, where no or ground-truth is available; (ii) novel self- supervision provides reliable predictions under challenging setting. The proposed uses neural model predict image, utilizes an optical validate refine prediction. verify three benchmark datasets with rendered stacks stacks. Qualitative quantitative evaluations method strong baseline supervised tasks. source code publicly avail- able at https://github.com/Ehzoahis/DEReD.

Language: Английский

---

Spatially varying defocus map estimation from a single image based on spatial aliasing sampling method

Optics Express, Journal Year: 2024, Volume and Issue: 32(6), P. 8959 - 8959

Published: Feb. 8, 2024

In current optical systems, defocus blur is inevitable due to the constrained depth of field. However, it difficult accurately identify amount at each pixel position as point spread function changes spatially. this paper, we introduce a histogram-invariant spatial aliasing sampling method for reconstructing all-in-focus images, which addresses challenge insufficient pixel-level annotated samples, and subsequently introduces high-resolution network estimating spatially varying maps from single image. The accuracy proposed evaluated on various synthetic real data. experimental results demonstrate that our model outperforms state-of-the-art methods map estimation significantly.

Language: Английский

---

2HDED:Net for Joint Depth Estimation and Image Deblurring from a Single Out-of-Focus Image

2022 IEEE International Conference on Image Processing (ICIP), Journal Year: 2022, Volume and Issue: unknown, P. 2006 - 2010

Published: Oct. 16, 2022

Depth estimation and all-in-focus image restoration from defocused RGB images are related problems, although most of the existing methods address them separately. The few approaches that solve both problems use a pipeline processing to derive depth or defocus map as an intermediary product serves support for deblurring, which remains primary goal. In this paper, we propose new Deep Neural Network (DNN) architecture performs in parallel tasks by attaching same importance. Our Two-headed Estimation Deblurring (2HDED:NET) is encoder-decoder network Defocus (DFD) extended with deblurring branch, sharing encoder. tested on NYU-Depth V2 dataset compared several state-of-the-art deblurring.

Language: Английский

---

HI-Net: Boosting Self-Supervised Indoor Depth Estimation via Pose Optimization

IEEE Robotics and Automation Letters, Journal Year: 2022, Volume and Issue: 8(1), P. 224 - 231

Published: Nov. 24, 2022

Pose estimation plays a critical role in self-supervised monocular depth for indoor scenes, especially those involving complex ego-motion. In this letter, we leverage the two-view geometry constraints into pose to boost accuracy of estimation, which ultimately improves performance estimation. Specifically, decompose two steps: initial homography and iterative residual refinement. We first introduce Homography Estimation Module (HEM) estimate large 3-DoF rotations. Then, refine 6-DoF with an Iterative Residual Refinement (IRM). Finally, supervision signal is generated refined used training DepthNet. Experiments on NYU V2 dataset show that our approach significantly DepthNet, proposed method achieves state-of-the-art results. Furthermore, experiments ScanNet demonstrate generalization ability both

Language: Английский

---

Self-Supervised Spatially Variant PSF Estimation for Aberration-Aware Depth-from-Defocus

ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Journal Year: 2024, Volume and Issue: unknown, P. 2560 - 2564

Published: March 18, 2024

In this paper, we address the task of aberration-aware depth-from- defocus (DfD), which takes account spatially variant point spread functions (PSFs) a real camera. To effectively obtain PSFs camera without requiring any ground-truth PSFs, propose novel self-supervised learning method that leverages pair sharp and blurred images, can be easily captured by changing aperture setting our PSF estimation, assume rotationally symmetric introduce polar coordinate system to more accurately learn estimation network. We also handle focus breathing phenomenon occurs in DfD situations. Experimental results on synthetic data demonstrate effectiveness regarding both depth estimation.

Language: Английский

---

Camera-Independent Single Image Depth Estimation from Defocus Blur

2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Journal Year: 2024, Volume and Issue: unknown, P. 3737 - 3746

Published: Jan. 3, 2024

Monocular depth estimation is an important step in many downstream tasks machine vision. We address the topic of estimating monocular from defocus blur which can yield more accurate results than semantic based methods. The existing techniques are sensitive to particular camera that images taken from. show how several camera-related parameters affect using optical physics equations and they make depend on these parameters. simple correction procedure we propose alleviate this problem does not require any retraining original model. created a synthetic dataset be used test independent performance models. evaluate our model both real datasets (DDFF12 NYU V2) obtained with different cameras methods significantly robust changes cameras. Code: https://github.com/sleekEagle/defocus_camind.git

Language: Английский

---

Are Realistic Training Data Necessary for Depth-from-Defocus Networks?

IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society, Journal Year: 2022, Volume and Issue: unknown, P. 1 - 6

Published: Oct. 17, 2022

Image-based depth estimation is one of the important tasks in computer vision. Depth-from-defocus (DfD) methods estimate scene from a single or multiple defocused images by exploiting depth-dependent defocus blur cues. Because difficulty obtaining real-world dataset with ground-truth depth, most deep-learning-based DfD rely on synthetic training dataset, where more realistic rendering considered desirable for accurate estimation. In this paper, we consider if 3D objects are really necessary networks. To investigate this, design very simple and fast data generation method using only two front-parallel texture planes compare it widely-applied path-tracing common object dataset. Through experiments, show that 2-plane provides comparable even slightly better performance than can be as an alternative practical network training.

Language: Английский

---

Measuring Focus Quality in Vector Valued Images for Shape from Focus

2022 26th International Conference on Pattern Recognition (ICPR), Journal Year: 2022, Volume and Issue: unknown

Published: Aug. 21, 2022

In shape from focus (SFF) methods, the measure (FM) operator plays a key role in determining ultimate of object. Usually, vector-valued (color) images are converted into grayscale before applying FM operator. This conversion saves computations; however, it affects accuracy values which deteriorates depth map. paper proposes an effective to find relative degree for pixels image sequence. first step, transformed scalar-valued by computing scaled norm resultant vector differences. The scaling factor is computed through various features based on operations including dot product, cross projections, and vectors distances. Then differential kernels with gap applied scalar compute values. Experiments conducted using synthetic real sequences reveal that proposed method providing better quality 3D shapes objects.

Language: Английский

---