2D-MobileStereoNet: Efficient Stereo Matching

• The paper introduces 2D-MobileStereoNet, which replaces heavyweight 3D cost volume aggregation with a novel 2D interlacing approach to achieve efficient stereo matching.
• It strategically uses MobileNet-V1 and MobileNet-V2 blocks with depthwise-separable convolutions to drastically reduce parameters and MACs while maintaining competitive accuracy.
• Ablation studies confirm that the interlacing cost volume construction lowers endpoint error and enhances performance on benchmarks like SceneFlow and KITTI.

2D-MobileStereoNet is a lightweight deep stereo matching architecture designed to achieve a balance between computational efficiency and accuracy, targeting deployment on resource-limited devices such as mobile GPUs. The core design replaces heavyweight 3D cost volume aggregation with a pure 2D convolutional approach, employing depthwise-separable MobileNet blocks and a learnable “interlacing” cost volume, yielding significant reductions in parameters and multiply–accumulate operations (MACs) while remaining competitive with state-of-the-art 3D architectures (Shamsafar et al., 2021).

1. Architectural Overview

2D-MobileStereoNet consists of the following sequential modules:

• Feature Extraction: An initial ResNet-style shared 2D backbone generates a 320 × (H/4) × (W/4) feature map from each stereo input. The backbone uses depthwise-separable MobileNet blocks, with the first three $3\times3$ convolutions replaced by MobileNet-V2 “inverted residual” blocks (expansion factor $t=3$), and all residual block convolutions switched to MobileNet-V1 blocks.
• Channel Reduction: Four consecutive $1\times1$ convolutions reduce the channel count from 320 to 32.
• Interlacing Cost Volume Construction: Instead of fixed correlation or naïve concatenation, left and right features are interlaced channel-wise at each disparity and passed through a small learnable 3D-conv subnetwork, parameterized by the grouping hyperparameter $i=4$, producing a $(D_{\max}/4) \times (H/4) \times (W/4)$ cost volume.
• 2D Encoder–Decoder (Hourglass): A deeply stacked, U-shaped 2D network (three hourglass modules, each with MobileNet-V2 blocks of width 48 channels) carries out cost aggregation.
• Disparity Regression: The network directly regresses the disparity map using the final hourglass output, upsampling to the input resolution.

2. Interlacing Cost Volume Construction

2D-MobileStereoNet introduces a cost volume construction mechanism tailored for 2D networks. For each disparity $d$, the right feature tensor $f_R$ is horizontally shifted by $d$; the left and shifted-right features are interlaced channel-wise, yielding an interlaced tensor. This tensor is processed by three 3D convolutions (kernels of size $(2i \times 3 \times 3)$) and then projected to a 1-channel cost by a 2D $3 \times 3$ convolution. The resulting cost volume feeds into the encoder–decoder aggregation network.

This “interlacing” approach outperforms conventional correlation and concatenation strategies in terms of endpoint error (EPE) on synthetic and real datasets, lowering the SceneFlow EPE to 1.55 px from 1.71 (correlation) and 1.86 (concat) (Shamsafar et al., 2021).

3. MobileNet Block Substitution

The architecture leverages two major variants of MobileNet blocks:

• MobileNet-V1 (v1) block: Applies a depthwise $t=3$0 convolution over $t=3$1 channels, followed by a $t=3$2 pointwise conv. This reduces MACs by $t=3$3 for $t=3$4, $t=3$5.
• MobileNet-V2 (v2) block (inverted residual): Expands channels by $t=3$6 via $t=3$7 conv, applies a $t=3$8 depthwise conv, then projects back with $t=3$9 conv, optionally with skip connections.

Key placements include: v2 ($1\times1$0) in the earliest backbone layers, v1 in backbone residuals, v2 ($1\times1$1) in pre-hourglass, v2 ($1\times1$2) in hourglass modules. This strategic substitution maintains accuracy (EPE 1.50 px on SceneFlow) and compresses parameter/MAC budgets.

4. Disparity Regression and Loss Function

The output tensor from the stacked hourglass is a single-channel disparity map $1\times1$3, upsampled to $1\times1$4 via bilinear interpolation. Disparity supervision employs a smooth L1 regression loss:

$1\times1$5

No explicit soft-argmax is used; the model operates as a direct regressor, which empirical results show can match 3D-aggregation baselines in accuracy (Shamsafar et al., 2021).

5. Computational Complexity and Benchmark Performance

The efficiency characteristics and accuracy are summarized in the following table:

Model	Params (M)	MACs (G)	SceneFlow EPE	KITTI15 D1_all (%)	Inference Footprint
DispNet-C	38	—	1.67	—	—
PSMNet (3D)	5.2	256	0.88	2.10	—
GwcNet-g (3D)	6.4	246	0.62	1.53	—
2D-MobileStereoNet	2.23	30	1.14	2.83 (test)	10 MB

2D-MobileStereoNet achieves 2.83% D1-all error on the KITTI 2015 test benchmark (27% fewer parameters and 95% fewer MACs than preexisting 3D networks). Model size (10 MB) and MACs (30 G) enable feasible deployment on consumer mobile and embedded hardware (Shamsafar et al., 2021).

6. Ablation Studies and Design Tradeoffs

Ablations conducted in (Shamsafar et al., 2021) establish:

• “Interlacing” cost volume construction consistently improves accuracy over correlation and plain concatenation, at minimal additional cost.
• MobileNet block substitution (v1 and v2 with appropriately tuned expansion $1\times1$6) leads to a significant reduction in parameters and MACs with negligible performance loss.
• The channel grouping parameter $1\times1$7 for cost volume construction was found optimal at $1\times1$8, balancing cost and accuracy.

Expansion $1\times1$9 set higher than recommended increases both error and compute cost, corroborating design choices that favor moderate expansion for early/low-level layers and more reserved expansion for dense prediction stages.

7. Comparative Analysis to BANet-2D and Limitations

BANet-2D subsequently extended the 2D-MobileStereoNet approach by introducing a scale-aware spatial attention mechanism that splits cost aggregation into “detailed” and “smooth” streams, each processed separately before fusion. This design, alongside the use of similar MobileNet-V2 backbones, yields substantially higher accuracy (1.83% D1-all on KITTI 2015) and significantly less latency (BANet-2D achieves ≈45 ms runtime, MobileStereoNet-2D ≈140 ms at 512×512 input, with ≈4× fewer FLOPs) (Xu et al., 5 Mar 2025). The findings indicate that while 2D-MobileStereoNet provides an efficient baseline, a single 2D aggregation stream can suffer from edge blurring and loss of structural detail, whereas more sophisticated split-aggregation frameworks offer a marked improvement on both accuracy and efficiency.

A plausible implication is that 2D-MobileStereoNet, in its original design, may not optimally handle fine-grained edge features or large smooth regions in challenging datasets. However, it remains one of the first competitive 2D-convolutional stereo approaches demonstrating that resource-efficient cost aggregation need not preclude state-of-the-art accuracy in practical stereo applications (Shamsafar et al., 2021).