DBStereo: Decoupled 4D Volume with 2D Aggregation

• DBStereo is a stereo matching network that decouples the 4D cost volume into separate spatial and disparity pathways using pure 2D convolutions.
• It leverages a Channel2Disp reshape and a Bidirectional Geometry Aggregation block to enforce spatial local smoothness and disparity unimodality.
• Empirical results show DBStereo achieves state-of-the-art real-time performance with lower memory and computational overhead compared to 3D aggregation methods.

Searching arXiv for DBStereo and closely related stereo-matching papers to ground the article in current literature. DBStereo is a stereo matching network introduced in “Decoupling Bidirectional Geometric Representations of 4D cost volume with 2D convolution” (Wei et al., 2 Sep 2025). It is a deployment-friendly 4D cost aggregation network based on pure 2D convolutions, designed to retain the representational richness of a full 4D cost volume while removing 3D convolutions from cost aggregation. The method is organized around a decoupling of the 4D cost volume into spatial geometry over $(H,W)$ and disparity geometry over $D$, implemented through a Channel2Disp reshape and a Bidirectional Geometry Aggregation block. Within contemporary stereo matching, DBStereo is positioned as an alternative to both 3D cost-volume regularization and iterative refinement schemes, with an emphasis on real-time performance and mobile-friendly deployment (Wei et al., 2 Sep 2025).

1. Problem setting and motivation

DBStereo addresses the standard learning-based stereo pipeline in which features are extracted from left and right images, a 4D cost volume is constructed, the volume is aggregated or regularized, and disparity is regressed. In the notation used for the method, a 4D cost volume has shape

$$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$

The established high-accuracy design is 3D cost-volume regularization with 3D CNNs, as in architectures such as GC-Net, PSMNet, GwcNet, and ACVNet. DBStereo is motivated by the observation that this design is memory-intensive, computationally expensive, and latency-heavy, particularly for mobile or edge deployment (Wei et al., 2 Sep 2025).

The paper contrasts this regime with two other families. First, 2D aggregation-based methods avoid 3D convolutions but typically lose robustness in ill-posed regions such as occlusions, textureless areas, repetitive patterns, and reflective surfaces. Second, iterative methods such as RAFT-Stereo and IGEV-Stereo reduce dependence on explicit 3D aggregation, yet can exhibit disparity discontinuities and artifacts in occlusion or low-texture regions and often require runtimes above 300 ms on large GPUs. DBStereo is therefore framed as an attempt to preserve the semantic richness of a 4D cost volume while making aggregation efficient enough for real-time use and simple enough for deployment with standard 2D operators (Wei et al., 2 Sep 2025).

A central claim of the method is that standard 3D convolution entangles spatial and disparity structure. DBStereo argues instead for two explicit priors: a spatial local smoothness prior, under which nearby pixels at the same depth tend to share similar disparity, and a disparity unimodality prior, under which the disparity distribution at each pixel should ideally form a single sharp mode. This decomposition provides the conceptual basis for the network’s decoupled aggregation design.

2. Decoupled bidirectional geometric representations

The defining idea of DBStereo is the decoupling of the 4D cost volume into two directions of geometry: spatial geometry over $(H,W)$ and disparity geometry over $D$ (Wei et al., 2 Sep 2025). Rather than regularizing the volume with kernels such as $3 \times 3 \times 3$ over $(D,H,W)$, the method reshapes disparity into a channel-like axis and then applies pure 2D convolutions.

The cost volume is constructed as a group-wise correlation volume:

$$\mathbf{C}_{gwc}(d, x, y, g) = \frac{1}{N_c/N_g} \left\langle \mathbf{f}_l^g(x,y), \mathbf{f}_r^g(x - d, y) \right\rangle,$$

which yields

$$\mathbf{C}_{gwc} \in \mathbb{R}^{D \times G \times H \times W}.$$

DBStereo then applies the Channel2Disp operator:

$$\mathbf{C}_{3D} = \text{Reshape}(\mathbf{C}_{gwc}) \in \mathbb{R}^{(G \cdot D) \times H \times W}.$$

This reshape fuses the disparity dimension $D$0 and group dimension $D$1 into a single channel axis. Spatial layout is preserved, but disparity is now encoded across channels rather than as an explicit third spatial dimension (Wei et al., 2 Sep 2025).

This representation enables two distinct forms of aggregation. Spatial aggregation uses standard 2D convolutions with spatial kernels such as $D$2 over $D$3, enforcing the local smoothness prior. Disparity aggregation uses $D$4 convolutions across the channel dimension; because channels encode disparity candidates, this implements global mixing over all disparity hypotheses at each pixel. In the paper’s interpretation, this global channel mixing is better aligned with disparity unimodality than local 3D kernels, whose receptive field over disparity grows only gradually with depth (Wei et al., 2 Sep 2025).

The term “bidirectional geometry aggregation” refers to the alternation of these two operations. Spatial aggregation refines geometry across neighboring pixels while preserving disparity-specific structure; disparity aggregation refines the distribution over disparity candidates at a fixed pixel. The alternation of the two creates a bidirectional interaction in which spatial context influences disparity selection and refined disparity distributions feed back into later spatial aggregation.

3. Network architecture

DBStereo follows a conventional stereo backbone up to cost construction, but replaces conventional 3D aggregation with a specialized Bidirectional Geometry Aggregation block. The feature extractor is MobileNetV2, pretrained on ImageNet, and it produces multi-scale left and right features

$D$5

Upsampling blocks restore features to $D$6 resolution, and the final multi-scale features are

$D$7

The pair $D$8 is used to build the 4D group-wise correlation cost volume, while $D$9 are used to generate a spatial attention map intended to enhance disparity robustness (Wei et al., 2 Sep 2025).

The core aggregation unit is the Bidirectional Geometry Aggregation, or BGA, block. It is described as an encoder–decoder architecture composed entirely of 2D operations. Its two constituent modules are:

• Spatial Aggregation Module: operates on $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$0 with 2D spatial convolutions, aggregating context across neighboring pixels while treating disparity channels as feature channels.
• Disparity Aggregation Module: also operates on $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$1 but uses $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$2 convolutions so that, at each pixel, all disparity channels interact through a fully connected channel mixing matrix.

The aggregation sequence is described conceptually as

$$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$3

No 3D convolutions are used anywhere in this stage. This is the architectural point on which the paper’s broader claim rests: the 4D cost-volume paradigm can be retained while aggregation is reformulated as a decoupled 2D problem (Wei et al., 2 Sep 2025).

After aggregation, disparity is regressed by softmax over disparity and soft-argmin:

$$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$4

The resulting low-resolution disparity map is then upsampled to full resolution using interpolation and learnable refinement to obtain the final disparity prediction.

4. Optimization, training protocol, and empirical results

DBStereo is trained with a standard smooth $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$5 loss on two supervision points: an initial or low-resolution disparity prediction and the final refined disparity. The total loss is

$$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$6

with $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$7 and $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$8 (Wei et al., 2 Sep 2025).

The training protocol uses PyTorch on 8 RTX 3090 GPUs. Pretraining is performed on Scene Flow, with 35,454 training pairs and 4,370 test pairs at 960×540 resolution, for 90 epochs. Fine-tuning is then performed on the combined KITTI 2012 and KITTI 2015 training sets for 500 epochs. KITTI 2012 contributes 194 train and 195 test pairs, while KITTI 2015 contributes 200 train and 200 test pairs (Wei et al., 2 Sep 2025).

The paper reports a family of DBStereo variants on Scene Flow, measured in EPE, D1, and runtime on RTX 3090:

Variant	EPE	D1	Runtime
DBStereo-S	0.65	2.36%	~15 ms
DBStereo-M	0.50	1.80%	~33 ms
DBStereo-L	0.45	1.57%	~49 ms

These figures are used to compare DBStereo against both aggregation-based and iterative stereo methods. On the same table, PSMNet reports EPE 1.09 with runtime ~317 ms, RAFT-Stereo reports EPE 0.61 and D1 2.85% with runtime ~380 ms, and IGEV-Stereo reports EPE 0.47 and D1 2.47% with runtime ~340 ms. DBStereo-L is therefore reported as better than IGEV-Stereo in EPE while being substantially faster, and as outperforming the listed 3D aggregation methods in both accuracy and runtime on that comparison (Wei et al., 2 Sep 2025).

KITTI 2012 and KITTI 2015 results are described qualitatively rather than fully tabulated in the provided material. The paper states that DBStereo reaches state-of-the-art performance among real-time methods and performs strongly on the online leaderboards. The qualitative interpretation attached to these results is that pure 2D convolutions can be sufficient for stereo cost aggregation if the geometry of the 4D volume is modeled correctly.

5. Efficiency, deployment, and limitations

The efficiency argument for DBStereo is architectural rather than merely empirical. By replacing 3D convolutions with standard 2D convolutions, including $$C_{4D} \in \mathbb{R}^{D \times C \times H \times W}.$$9 spatial kernels and $(H,W)$0 channel-mixing kernels, the method uses operators that are already highly optimized on commodity hardware and mobile accelerators (Wei et al., 2 Sep 2025). The paper explicitly notes several consequences: lower FLOPs and memory than 3D aggregation networks at similar accuracy, better cache locality, and easier deployment through systems that are friendly to 2D convolutional graphs.

The same section provides a complexity comparison in which DBStereo reports 62.65 GFLOPs and 19.6 M parameters. For context, MonoDETR reports 62.96 GFLOPs and 37.7 M parameters, YOLOStereo3D reports 177.82 GFLOPs and 107.6 M parameters, and S3AD reports 180.71 GFLOPs and 109.5 M parameters; however, those latter methods are 3D object detectors rather than stereo matchers, so the comparison is informative primarily as an indicator of the computational scale of the 2D-only design (Mu et al., 23 Jun 2026). Within stereo matching itself, the paper’s main deployment claim is that pure 2D convolution is more naturally amenable to mobile accelerators, TensorRT, NNAPI, and CoreML-style deployment targets (Wei et al., 2 Sep 2025).

The method’s limitations are also explicit. DBStereo still relies on a full 4D cost volume, which can remain memory-heavy at high resolutions and large disparity ranges. After Channel2Disp, the channel dimension $(H,W)$1 can become large, so although $(H,W)$2 convolutions are efficient, extreme disparity ranges may still be challenging on low-end devices. The paper also identifies further optimization opportunities through pruning, quantization, and more sophisticated multi-scale strategies (Wei et al., 2 Sep 2025).

A plausible implication is that DBStereo does not remove the representational cost of stereo matching so much as shift its computational bottleneck into a form that current deployment stacks handle better. This interpretation is consistent with the paper’s repeated emphasis on deployment friendliness rather than on eliminating the cost-volume paradigm itself.

6. Position within stereo research and terminological boundaries

DBStereo is best understood as a reformulation of 4D cost aggregation rather than a departure from cost-volume stereo. It preserves the group-wise correlation volume and the soft-argmin disparity regression familiar from earlier stereo systems, but rejects the assumption that 4D volumes must be regularized with 3D CNNs. In that sense, it occupies a middle position between heavy 3D aggregation methods and iterative correlation-based approaches. The paper explicitly presents it as a challenge to the empirical design rule that 4D cost volumes require 3D convolutions (Wei et al., 2 Sep 2025).

Its broader significance lies in the form of inductive bias it introduces. The separation between spatial local smoothness and disparity unimodality is not merely a computational device; it is proposed as a geometry-aware decomposition of the stereo problem. This suggests a wider “decoupled aggregation” paradigm for other high-dimensional matching tasks such as optical flow, multi-view stereo, or depth from video, all of which are named as future directions in the paper (Wei et al., 2 Sep 2025).

The name should also be distinguished from similarly named methods in adjacent subfields. DDStereo, for example, is a distinct stereo Transformer for real-time open-set 3D object detection and is not the same method (Mu et al., 23 Jun 2026). StereoISP, by contrast, uses stereo disparity in the raw sensor domain for demosaicking and denoising rather than for disparity estimation itself (Rabiah et al., 2022). These neighboring usages illustrate that “stereo-aware” design can refer to different levels of the vision pipeline, whereas DBStereo in the strict sense designates a stereo matching network centered on decoupled 2D aggregation of a 4D cost volume.

Within the stereo matching literature, DBStereo is therefore notable for three tightly connected reasons: it retains the full 4D cost-volume formulation, replaces 3D regularization with pure 2D convolution, and encodes spatial and disparity geometry as separate but alternating aggregation directions. The paper’s own conclusion is that this combination yields a simple yet strong baseline for further study of decoupled aggregation in stereo matching (Wei et al., 2 Sep 2025).