2.1 Cost volume construction and aggregation

Constructing cost volume is a core aspect of the stereo matching task, which directly affects the accuracy and computational efficiency of the disparity estimation. Recently, significant progress has been made in constructing the cost volume using approaches based on learning. Comprehensive image feature extraction is a key step in building an effective cost volume, and the effectiveness of multi-scale mechanisms in capturing contextual information has been widely validated across multiple domains. For example, ESTINet [23] incorporates a multi-scale strategy in its Spatial Information Collection Module (SICM), where features at different spatial resolutions are extracted and fused to improve the overall quality of video deraining. In the single-image desnowing domain, multi-scale mechanisms have likewise seen broad adoption. DDMSNet [24] introduces multi-scale designs in both the pixel and feature domains, enabling the simultaneous capture of fine-grained textures and global context, thereby further enhancing desnowing restoration performance. GridFormer [25] proposes a grid-structured multi-scale interaction module for image restoration under adverse weather. this module balances global dependencies and local details while effectively controlling computational cost, achieving superior performance and exhibiting high scalability. In addition, MB-TaylorFormer V2 [26] introduces a multi-scale patch embedding module that employs multi-branch architectures, deformable convolutions, and convolutions with different receptive fields to extract feature information, thereby providing information-rich and flexible tokens for subsequent processing.

In the field of stereo matching, PSMNet [6] constructs cost volume at different scales by introducing spatial pyramid pooling and uses stacked hourglass 3D convolutional neural networks to supervise the specification of the cost volume, which is a milestone work. Subsequent CFNet [27] builds upon PSMNet [6] and introduces a cascade cost volume fusion-based network. This network fuses multiple low-resolution cost volumes to generate a feature-rich cost volume, thereby enhancing the robustness of the stereo matching network. However, the cubic computational complexity and high memory consumption of 3D CNNs are very expensive to deploy in practical applications, for this reason, AANet [10] proposed a cross-scale cost aggregation method via sparse point sets and a neural network-based cross-scale cost aggregation algorithm, which efficiently solves the edge diffusion problem and handles texture-free regions. IGEV-Stereo [28] constructs a combined geometric codomain that encodes geometric features and contextual information, and this network can be used to deal with locally blurred regions in images by effectively capturing non-local geometric information. Mocha-Stereo [29] proposed a Motif channel correlation volume for solving the problem of mismatched edge details due to the loss of geometric structural information in the feature channel generation process. DFMNet [30] introduces a cost-filtering volume that leverages guidance weights derived from group-wise correlations to filter redundant information in the concatenated volume, thereby improving the accuracy of feature representations. LoS [31] proposes a novel local structure representation, termed Local Structure Information (LSI), to replace the 3D cost volume, enabling more accurate stereo matching without the need to construct a 3D cost volume. IVF-Stereo [32] addresses the issue of cost volume distortion in asymmetric stereo matching by iteratively fusing complementary correlation volumes and concatenated volumes, combined with multi-peak search and two-stage optimization.

Some researchers introduced the attention mechanism into the stereo matching task, STTR [33] revisited the stereo depth estimation problem from the perspective of sequence-to-sequence correspondence by adding Transformer [34] to a stereo matching network for dense pixel matching alternative cost volume building using positional information and the attention mechanism. ACVNet [35] designed an innovative approach for cost volume construction that uses correlation clues to generate attention weights, suppressing redundant information and enhancing matching-related information. Additionally, it employs multi-layer adaptive patch matching to improve the distinctiveness of the matching cost at different disparity levels.

There are many works devoted to study how to efficiently cost aggregation, CSPN [36] and GANet [37] use spatial transformations to aggregate costs. CSPN [36] learns the affinity of neighbouring pixels through a deep convolutional network to form an affinity matrix, which performs the cost aggregation. GANet [37] proposed semi-global and local guided aggregation layers to replace computationally expensive 3D convolutional layers, thereby achieving efficient and accurate cost aggregation. These methods achieve high accuracy depth estimation at the expense of more time. StereoNet [38], ADCPNet [39] and IINet [40] focus on building real-time stereo matching networks, but these networks often achieve fast depth estimation by sacrificing some accuracy. We use 3D convolution-based spatial filtering, but focuses on both speed and efficiency.

2.2 Disparity refinement

Disparity refinement is a key post-processing aspect in stereo matching algorithms, aiming to improve the accuracy and robustness of the initial disparity estimation. In recent years, methods based on iterative optimisation have achieved excellent performance in stereo matching work. Zachary Teed firstly proposed the iterative model RAFT [41], which constructs a 4D cost volume by calculating the relationship between all pixel pairs, and then retrieves iterative features from the correlation volume for optical flow estimation using a GRU-based update operator. Later, Lahav Lipson migrated the RAFT framework to stereo matching by constructing the RAFT-Stereo [16] stereo matching network, which introduces multiple levels of GRU convolution to iteratively update the disparity values by using local surrogate values retrieved from the correlations of all the pixel pairs to improve the accuracy of the disparity estimation. CREStereo [42] improved on RAFT-Stereo by introducing an adaptive group correlation module to improve the generalisation of the network, as well as a recursive updating module and a cascade stacking module, which are used to obtain finer disparity maps. Many stereo matching works proposed specific modules to deal with the initial disparity in order to improve the accuracy of the disparity maps. StereoDRNet [43] proposed a novel disparity refinement network that uses geometric error as input for optimization, effectively addressing the issue of inconsistency in geometric information in the disparity map. Eai-Stereo [44], in response to the blurriness and lack of detail in the generated disparity maps, proposed an error-aware module, which combines the high-frequency information of the original image to improve the error correction capability of the network, thus generating excellent disparity maps. DLNR [18] proposed a disparity normalization refinement module, which, by normalizing the disparity with respect to the image width, addresses the issue of module failure in cross-domain scenarios.

In recent years, frequency domain decomposition has been widely applied in the field of computer vision. HLNet [45] designs a high-low frequency decomposition-based approach for image restoration and enhancement. By introducing high-low frequency decomposition blocks, it can effectively handle different types of image degradation, thereby achieving higher-quality image restoration. HiLoF-GAN [46] is a novel underwater image enhancement method that leverages Generative Adversarial Network (GAN) combined with high-low frequency separation, effectively improving the quality and visual effects of underwater images. In the field of video anomaly detection, FE-VAD [47] incorporates high-low frequency enhancement modules to separate and modulate high-low frequency information in both spatial and temporal domains. We will use a two-stage disparity refinement in order to further improve the accuracy of the disparity map, with the first stage being a multi-scale GRU-based disparity refinement and the second stage being a frequency domain decomposition-based disparity reconstruction.

3.1 Feature extraction

Feature extraction consists of two parts: 1) a multi-scale feature extraction network that extracts multi-scale features for constructing cost volumes and guiding cost volume aggregation, and 2) a multi-scale context extraction network that extracts multi-scale context features for iterative updating of disparity based on multi-scale GRUs.

Multi-scale Feature Extraction Network:for the corrected left and right images, we first use the MobileNetV2 [48] model pre-trained on ImageNet [49] to extract features, producing feature maps at 1/32 of the original image resolution. We then use an upsampling module with skip connections to progressively restore the feature maps to resolutions of 1/16, 1/8, and 1/4 of the original image, thereby obtaining multi-scale feature maps. We use the 1/4 resolution feature map to construct the cost volume, and 1/4, 1/8, 1/16, and 1/32 resolution images are used to guide the cost aggregation.

Multi-scale Context Extraction Network: this network is applied to the left image only and consists of a series of downsampling layers and residual blocks that generate feature maps at 1/4, 1/8, 1/16 of the input image resolution, each with 128 channels. Contextual features are used to initialise the hidden state of the multi-scale GRU, and during the iterative refinement of disparity, the GRU generates updated hidden states.

3.2 Cost volume integrating geometric information

3.2.1 Cost volume construction.

We construct the initial cost volume using the 1/4 resolution feature map of the original image, extracted by the multi-scale feature extraction network. We construct a grouped correlation volume using the GwcNet method, which divides the features into groups along the channel dimensions and calculates the correlation volume group by group.

(1)

Where N_c is the number of feature channels, d is the index of disparity, and is the inner product.

is limited by local similarity due to its dependence on feature correlation and cannot adequately express the global geometric information of an image. Inspired by the fact that 3D convolution can effectively encode geometric information, we propose a Multi-scale Geometric Information (MGI) module that achieves an effective transformation from local correlation to global geometric understanding. The module is an encoder-decoder structure, where the encoder and decoder consist of three downsampling blocks and three upsampling blocks, respectively, with each downsampling block consists of two 3×3×3 convolutions, and each upsampling block consists of a 3D transpose convolution with a kernel size of 4×4×4. is downsampled to 1/2, 1/4, and 1/8 of the original resolution by the encoder, with the cost volume having 16, 32, and 48 channels at each respective downsampling stage. In the decoder stage, the cost volume is gradually upsampled by the transpose convolution to restore to the original resolution. In order to achieve the effective fusion of multi-scale features, we also introduce a skip connection mechanism to merge the output results of the encoder and into the decoder stage, which can enable the network to simultaneously utilise both shallow and deep features, and to capture both detailed features and deep semantic information, in order to enhance the expressive power of the geometric features of the geometric cost volume.

3.2.2 Feature channel attention-guided cost aggregation.

The cost aggregation method using spatial transformation can effectively capture spatial neighbourhood information, but this method uses a large number of 3D convolutions, takes a long time to compute, and leads to increased memory consumption. For this reason, we introduce a lightweight Feature Channel Attention Guided (FCAG) mechanism, which uses the attention weights generated by the feature maps to dynamically adjust the importance of each channel in the cost volume. For a cost volume of size , feature maps of the same resolution are fed into the network, and channel attention weights are generated for each pixel of each channel, which are then applied to the cost volume to guide cost aggregation. We integrate FCAG into the cost aggregation to effectively improve the efficiency of aggregation stage.

(2)(3)

where σ is a sigmoid function and is realised by 2D convolution.

3.3 Disparity refinement

In the stage of disparity refinement, MGFD-Stereo is divided into two stages: the first stage is the iterative refinement of the disparity based on multi-scale GRU, and the second stage is the high and low-frequency error-separated disparity reconstruction.

3.3.1 GRU-based iterative refinement of disparity.

To obtain the initial disparity map, we use the soft max method to perform regression calculations from the cost volume , which can be expressed as:

(4)

where d is a set of discrete disparity indices.

In the iterative update phase of the multiscale GRU-based disparity, we predict a series of disparity fields starting from the initial disparity d₀ = 0. In each iterative update, a set of relevant features are indexed from a correlation pyramid based on the current value of the disparity d_i. The correlation pyramid here is a composite geometric pyramid constructed by combining relevant geometric features and relevant pixel features. The correlation features and disparity values are then processed through a encoding layer and concatenated with contextual features to obtain x_i, which is input to the multi-scale GRU.

(5)(6)(7)(8)(9)

Where x_i is the current input disparity value, z_i is the update gate, r_i resets the gate, and c_i, c_r, c_h and generate context features for the context feature network.

The hidden state h_i is obtained, which is decoded by the convolutional layer to obtain the disparity, which is used to update the disparity and can be expressed as:

(10)

3.3.2 High and low frequency separation for disparity refinement.

The iteratively updated disparity map has 1/4 resolution of the original image, which is up-sampled to get the full-resolution disparity map. In order to get a finer disparity map, we propose the High and Low Frequency Separation Disparity Reconstruction (HLFS). We use the error E as the input to this module. The error E is obtained by warping the right image to the left view based on the disparity and then computing the difference with the left image, which can be expressed as:

(11)

Where disp is the upsampled disparity map after iterative update, and warp is the warping transformation function.

As shown in Fig 3, we input the iteratively updated disparity map into a lightweight Encoder-Decoder (E-D) network to obtain multi-scale and multi-channel disparity maps. Our HLFS module is primarily divided into three branches. The upper branch is the High Frequency Information (HFI) branch, where we apply no convolution operations to preserve the high-frequency local detail information in the original high-resolution disparity map. The middle branch is the Adaptive Error Adjustment (AEA) branch, which generates an adaptive weight map through a series of convolution operations followed by an activation function, used to adjust the proportions of high- and low-frequency errors in the error map. The lower branch is the Low Frequency Information (LFI) branch, where adaptive average pooling serves as a low-frequency filter to extract low-frequency information from the image while attenuating high-frequency information to some extent. By employing HLFS for disparity refinement, we can obtain a full-resolution refined disparity map. The specific implementation of HLFS is represented as:

(12)(13)(14)

where is the Hadamard product, and is the disparity map before refinement.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. High and low-frequency separated disparity reconstruction module.

https://doi.org/10.1371/journal.pone.0340473.g003

3.4 Loss function

The overall loss of MGFD-Stereo consists of two parts. The first part is the initial disparity loss, which uses the Smooth L1 loss function to measure the error between the initial disparity and the ground truth disparity. Its expression is as follows:

(15)

where d₀ is the initial disparity value, d_gt is the ground truth disparity value, and the Smooth L1 loss is defined as follows:

(16)

The second part of the loss is the error between the multistage predicted disparity and the true disparity, which is expressed as follows:

(17)

Where N is the number of disparity iterations and . Therefore, the total loss function is expressed as follows:

(18)

4.1 Datasets and evaluation metrics

To test the performance of the proposed algorithm, we evaluate it on four publicly available datasets, Scene Flow [9], KITTI2012 [21], KITTI2015 [22], ETH3D [50] and Middlebury [51].

Scene Flow [9] is a synthetic stereo dataset containing 35,454 pairs of training samples and 4,370 pairs of testing samples, each with detailed dense disparity maps. SceneFlow includes two series of datasets: Cleanpass and Finalpass. This experiment uses the Finalpass dataset because it is closer to real-world images, making it more suitable for training and evaluating model performance. The experiment employs two metrics to assess the algorithm’s performance: the End-Point Error (EPE) and the three-pixel error percentage (D1).

KITTI 2012 [21] and KITTI 2015 [22] are datasets of real driving scenarios. KITTI 2012 includes 194 training image pairs and 195 testing image pairs, while KITTI 2015 includes 200 training image pairs and 200 testing image pairs. For the KITTI 2012 dataset, this experiment employs the percentage of erroneous pixels and the average end point error (EPE) in both all pixel regions (all) and non-occluded regions (noc) as evaluation metrics. For the KITTI 2015 dataset, the experiment adopts the three-pixel error percentages for background pixels (D1-bg), foreground pixels (D1-fg), and all pixels (D1-all), evaluated in both all pixel regions and non-occluded pixel regions, as evaluation metrics.

ETH3D [50] is a grayscale image dataset comprising 27 training image pairs and 20 testing image pairs. Middlebury 2014 [51] is an indoor scene dataset which consists of 15 sets of training image pairs and 15 sets of test image pairs, all of which are available in three different resolution formats. The ETH3D dataset uses the 1-pixel error as the experimental evaluation metric, while the Middlebury 2014 dataset uses the 2-pixel error as the experimental evaluation metric.

4.2 Implementation details

MGFD-Stereo is implemented using PyTorch framework and experimented on NVIDIA RTX 4090 GPU. In the experiments, this paper uses AdamW optimizer and crops the gradient to the range of [–1,1] for all the training datasets for brightness, colour saturation, scaling and flipping data. When training the MGFD-Stereo model on the Scene Flow dataset, the batch size is set to 8, the image is randomly cropped to 320 in height and 736 in width, the learning rate is 0.0002, and the training step size is 200k, and the model performance is monitored using the Scene Flow validation set. After that, the pre-trained model was fine-tuned on the KITTI2012 and KITTI2015 datasets by setting the batch size to 8, image random cropping height to 320, width to 736, learning rate to 0.0002, and training step size to 30k.

4.3 Ablation study

In order to better validate our proposed model structure, we conducted a large number of ablation experiments on the Scene Flow dataset, with all hyper-parameters set the same as the pre-training, and all the iteration rounds of the ablation experiments are 32, and the experimental results are shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Ablation experiments on the Scene Flow dataset.

https://doi.org/10.1371/journal.pone.0340473.t001

The experimental results show that, when no additional modules were incorporated, the baseline model achieved an EPE of 0.53 px and a D1 error rate of 2.78%. Firstly, the introduction of the MGI module reduces the EPE to 0.50px and the D1 error rate to 2.67%, indicating that the MGI enables the cost volume to capture the geometric information better and improves the accuracy of the disparity estimation. The FCAG module is then added to the baseline model, and compared to the baseline model, the EPE decreases by 7.55% and the D1 error rate decreases by 7.91%, which fully demonstrates the effectiveness of our proposed module. Finally, we added the MGI, FCAG, and FLFS modules to the baseline model at the same time, which led to the optimal performance of the baseline model, with the EPE and D1 error rates of 0.45px and 2.49%, respectively, demonstrating that our FLFS module is able to correct the disparity maps further, resulting in a more refined full-resolution disparity map. This series of experimental results not only validates the effectiveness of each proposed module, but also shows that there is a synergistic effect between them, which together form an end-to-end accurate stereo matching framework.

To validate the effectiveness of the high- and low-frequency branches in the HFLS module, we conducted ablation studies on the Scene Flow dataset. All hyperparameter settings for the ablation studies were identical to those used during pretraining. As shown in Table 2, we separately evaluated the performance of configurations without the HLFS module, with only the low-frequency branch (HLFS-LFI), with only the high-frequency branch (HLFS-HFI), and with the complete HLFS module. The experimental results show that when only LFI is retained, EPE does not decrease while D1 decreases slightly, indicating that low-frequency information plays a fundamental role in modeling the overall structure of the disparity map. Because LFI attenuates part of the high-frequency errors,the resulting performance improvement of the model is limited. When retaining only the HFI, the EPE is reduced by 0.02 px, indicating that high-frequency information plays an important role in refining details and correcting edges. High-frequency information typically captures local image details and is beneficial for both deblurring and denoising. Conversely, low-frequency information generally represents the global structure of the image, aiding the description of background and contours. Therefore, separating high- and low-frequency errors and applying targeted corrections can enhance the accuracy of disparity maps.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Ablation experiments of HLFS on the Scene Flow dataset.

https://doi.org/10.1371/journal.pone.0340473.t002

During the fine-tuning process, we designed a series of ablation experiments to verify the effects of different settings and to obtain optimal performance. Two fine-tuning strategies were adopted: (1) Mixed fine-tuning, where the pre-trained model was fine-tuned using a combined training set from KITTI 2012 and KITTI 2015, with training steps of 30k, 40k, and 50k; (2) Independent fine-tuning, where the pre-trained model was fine-tuned on a single dataset, either the KITTI 2012 or KITTI 2015 training set, also with training steps of 30k, 40k, and 50k. The quantitative evaluation results are presented in Tables 3 and 4. Overall, increasing the number of fine-tuning steps does not necessarily yield better performance, as excessively long training tends to degrade the generalization capability of the method. For example, in the KITTI 2015 test results, both mixed and independent fine-tuning achieved the best performance at 30k training steps, while performance declined at 40k and 50k steps. Furthermore, mixed fine-tuning generally outperformed independent fine-tuning. This can be attributed to the limited size of individual datasets, which increases the risk of model overfitting. KITTI 2012 mainly contains static scenes, while KITTI 2015 includes a large number of dynamic objects. Thus, the two datasets are complementary in terms of data distribution and scene characteristics. By employing mixed fine-tuning, the diversity of training samples is increased, thereby improving the robustness of the model. A comprehensive analysis of the experimental results indicates that employing mixed fine-tuning with 30k training steps yields excellent performance on both test sets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Quantitative evaluation of Independent fine-tuning (2015) and Mixed fine-tuning (2012/15) on KITTI 2015.

https://doi.org/10.1371/journal.pone.0340473.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Quantitative evaluation of Independent fine-tuning (2012) and Mixed fine-tuning (2012/15) on KITTI 2012.

https://doi.org/10.1371/journal.pone.0340473.t004

4.4 Comparisons with state-of-the-art

To further validate the performance of the method, we compare it with other state-of-the-art methods on the datasets Scene Flow, KITTI2012, KITTI2015, ETH3D and Middlebury.

Table 5 shows the performance comparison between our proposed method and existing advanced stereo matching methods on the Scene Flow dataset. As shown in the table, our method achieves an end-point error of 0.45px, representing error reductions of 58.7% and 40.8% compared to classic stereo matching methods PSMNet [6] and GwcNet [15], respectively. Compared to recent state-of-the-art methods such as RAFT-Stereo [16], ACVNet [35], IGEV-Stereo [28], and DLNR [18], our method still achieves certain performance improvements. In particular, it reduces the error by approximately 4.3% compared to the current state-of-the-art IGEV-Stereo. These experimental results convincingly demonstrate the effectiveness and robustness of our method on synthetic scene datasets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Comparison with state-of-the-art methods on the Scene Flow dataset.

https://doi.org/10.1371/journal.pone.0340473.t005

To validate the performance of our method in real driving scenarios, we conducted benchmark tests on the KITTI2015 and KITTI2012 datasets, with results shown in Tables 6 and 7. In the KITTI2015 stereo matching benchmark, our method was compared with state-of-the-art stereo matching approaches including PSMNet [6], GwcNet [15], CFNet [23], RAFT-Stereo [16], CREStereo [42], IGEV-Stereo [28], DFMNet [30], LoS [31] and DLNR [18]. Our method achieves D1-bg and D1-fg error rates of 1.39% and 2.54% respectively across all pixel regions, demonstrating outstanding performance. This indicates that our approach delivers highly accurate depth estimation in both static and dynamic areas of the scene, highlighting its robustness and generalization capability. Our method attains low error rates when processing both all-pixel regions and non-occluded pixel regions, with minimal difference between these error rates, indicating that the method exhibits robustness when handling occluded regions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Quantitative evaluation on KITTI 2015.

https://doi.org/10.1371/journal.pone.0340473.t006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Quantitative evaluation on KITTI 2012.

https://doi.org/10.1371/journal.pone.0340473.t007

In addition, we conducted quantitative comparisons with state-of-the-art methods including PSMNet [6], GwcNet [15], RAFT-Stereo [16], CREStereo [42], IGEV-Stereo [28], DFMNet [30], IVF-AStereo [32] and DVANet [52] on both the all regions and reflective regions of the KITTI 2012 test set. Experimental results show that while our method did not achieve optimal performance across all metrics in the all regions, it remains highly competitive compared to the state-of-the-art methods. However, our method achieves the best performance across all metrics in reflective regions, indicating superior robustness and accuracy when handling complex scenes and reflective regions. Our method focuses on edge details and complex scenes, while disparity optimization in regular regions is relatively limited. Therefore, the metrics in some regions are not optimal.

To further evaluate the generalization capability of our method, we compared its performance on the ETH3D and Middlebury datasets, with results shown in Table 8. Although our method did not achieve optimal results on the Middlebury and ETH3D datasets, it demonstrates excellent performance across different resolutions and scenes, showing competitive capability against advanced methods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Quantitative evaluation on Middlebury and ETH3D.

https://doi.org/10.1371/journal.pone.0340473.t008

4.5 Visualization results and analysis

In order to visually verify the effectiveness of the proposed method and deeply understand its performance in different scenarios, we conducted detailed visualisation results and analysis. In the datasets KITTI2012 and KITTI2015, we compare our method with state-of-the-art methods PSMNet [6], GwcNet [15], and CREStereo [42], as shown in Figs 4 and 5. The visualisation results show that our method is able to accurately capture the detailed regions in the scene, such as fences, parking stakes, and vehicle edges, which suggests that our method’s performance in matching edge details is excellent and that our method can accurately match the background and foreground regions of the scene to avoid confusion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Visualization Results on the KITTI2012 Dataset.

https://doi.org/10.1371/journal.pone.0340473.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Visualization Results on the KITTI2015 Dataset.

https://doi.org/10.1371/journal.pone.0340473.g005

In the dataset Middlebury, we compared with the state-of-the-art method RAFT-Stereo [16], as shown in Fig 6. Compared with RAFT-Stereo [16], our method is able to better preserve the detail information, effectively solving the problem of matching blurring at the edges and details, and the disparity maps estimated by our method are smoother and more coherent, with a significant improvement in detail prediction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Visualization Results on the Middlebury Dataset.

https://doi.org/10.1371/journal.pone.0340473.g006