BoRe-Depth: Boundary-Aware Monocular Depth

• BoRe-Depth is a self-supervised monocular depth estimation model that combines a lightweight encoder-decoder with explicit boundary refinement to enhance depth quality.
• It utilizes an Enhanced Feature Adaptive Fusion Module and pseudo-depth supervision to achieve accurate depth mapping with 8.7M parameters and 50.7 FPS on NVIDIA Jetson Orin.
• The two-stage training strategy incorporates semantic guidance after initial geometric learning, improving performance across benchmarks like NYUv2 and KITTI.

BoRe-Depth is a self-supervised monocular depth estimation model for embedded systems that combines a lightweight encoder–decoder architecture with explicit boundary refinement, semantic feature guidance, and pseudo-depth supervision in order to improve both numerical depth quality and object-boundary sharpness under strict runtime and parameter constraints (Liu et al., 6 Nov 2025). Introduced by Chang Liu, Juan Li, Sheng Zhang, Chang Liu, Jie Li, and Xu Zhang, the method is framed around four competing objectives—accuracy, boundary sharpness, model size, and real-time inference—and is explicitly reported as an 8.7M-parameter model running at 50.7 FPS on NVIDIA Jetson Orin (Liu et al., 6 Nov 2025). Within the cited literature, this is the explicit use of the name “BoRe-Depth”; in a separate borehole-imager depth-matching paper, “BoRe-Depth” does not appear as the formal method name and is described only as a possible external or project-level label for a ShapeDTW-based workflow (Li et al., 1 Dec 2025).

1. Scope and problem formulation

BoRe-Depth addresses self-supervised monocular depth estimation in the specific regime where embedded deployment and boundary fidelity are both first-order requirements (Liu et al., 6 Nov 2025). The motivating claim is that many lightweight self-supervised methods achieve efficiency by simplifying the decoder and relying on repetitive upsampling, but that this commonly produces oversmoothed predictions and blurred depth discontinuities. In the paper’s formulation, the practical failure mode is not limited to scalar depth error: boundary degradation can distort object shapes and fragment structures in the depth map or derived point cloud.

The model therefore treats “boundary refinement” as a central design problem rather than as a secondary visual quality issue. In this context, boundary refinement means improving predicted depth discontinuities so that object contours in the estimated depth map align more faithfully with scene structure. This is operationalized through decoder-side feature fusion, pseudo-depth boundary supervision, and semantic guidance. The target deployment domain is unmanned or embedded perception, where the use of monocular input is attractive because of its low-cost advantage, but where compute, memory, and runtime budgets are restrictive (Liu et al., 6 Nov 2025).

BoRe-Depth remains within the standard self-supervised monocular-video paradigm derived from view synthesis. It predicts depth from a target frame and camera motion from adjacent frames, then uses reconstruction and consistency losses to train without dense ground-truth depth. Its technical novelty lies not in abandoning this formulation, but in altering the architecture and supervision so that lightweight inference does not entail systematically degraded boundary quality (Liu et al., 6 Nov 2025).

2. Architecture and boundary-oriented feature fusion

The core prediction network is an encoder–decoder DepthNet coupled with a separate PoseNet for camera motion estimation (Liu et al., 6 Nov 2025). DepthNet uses MPViT, specifically MPViT-tiny in the lightweight configuration, as the encoder. For an input image

$I \in \mathbb{R}^{H \times W \times 3},$

the encoder produces five levels of depth features

$F^i,\quad i=1,2,3,4,5.$

The paper states that MPViT consists of a stem layer and four transformer encoders, with MPViT-tiny contributing 5.8M parameters to the overall design.

The decoder follows a hierarchical coarse-to-fine aggregation strategy. Its main architectural novelty is the Enhanced Feature Adaptive Fusion Module, or EFAF, which is designed to improve adjacent-scale fusion before detail recovery is lost through cheap upsampling. EFAF contains a Spatial Channel Enhancement block, defined as

$$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$

$$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$

where $DWConv_{3\times3}$ is depth-wise convolution, $PWConv_{1\times1}$ is point-wise convolution, and $\phi$ is GELU (Liu et al., 6 Nov 2025). The residual form preserves the original feature while injecting an enhanced local representation.

Fusion then proceeds hierarchically: $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$ with $\oplus$ denoting concatenation (Liu et al., 6 Nov 2025). The paper characterizes this as adaptive fusion, but the explicit equations are concatenation plus convolution rather than an explicit attention or gating law. Different levels do not share SCE weights, because the content distribution differs across scales.

This design suggests a decoder philosophy in which boundary preservation is achieved by strengthening each branch before multi-scale aggregation, rather than by relying on post hoc sharpening. The paper attributes the resulting gains to better use of coarse semantic context and fine spatial detail in adjacent-scale fusion, while maintaining a lightweight convolutional structure compatible with embedded deployment (Liu et al., 6 Nov 2025).

3. Training strategy, semantic guidance, and boundary supervision

BoRe-Depth uses a two-stage training strategy that combines self-supervised view synthesis with pseudo-depth boundary supervision and later semantic feature alignment (Liu et al., 6 Nov 2025). PoseNet predicts a 6D camera pose $P_t$ from adjacent frames such as $F^i,\quad i=1,2,3,4,5.$0 and $F^i,\quad i=1,2,3,4,5.$1, enabling standard reprojection-based supervision. The view reconstruction loss is

$F^i,\quad i=1,2,3,4,5.$2

with $F^i,\quad i=1,2,3,4,5.$3 in general (Liu et al., 6 Nov 2025). A geometric consistency term is also included: $F^i,\quad i=1,2,3,4,5.$4 where $F^i,\quad i=1,2,3,4,5.$5 is the predicted depth and $F^i,\quad i=1,2,3,4,5.$6 is the warped depth, although the exact form of $F^i,\quad i=1,2,3,4,5.$7 is not specified in the text.

Boundary supervision uses pseudo-depth labels $F^i,\quad i=1,2,3,4,5.$8 from a large monocular depth model. The paper states that these pseudo labels provide clearer boundaries than ground-truth depth labels, while also noting that they have limited accuracy and inherent model bias (Liu et al., 6 Nov 2025). Boundary alignment is enforced by

$F^i,\quad i=1,2,3,4,5.$9

where $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$0 denotes a normal or gradient operator on the depth map, $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$1 denotes a boundary operator, and the boundary is computed by a $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$2 Sobel operator. The balancing weights are usually equal and set to $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$3 (Liu et al., 6 Nov 2025). The exact similarity function $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$4 is not specified.

The first training stage uses

$$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$5

with the best setting reported as

$$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$6

The second stage adds semantic guidance: $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$7 with $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$8 (Liu et al., 6 Nov 2025).

The semantic loss aligns depth-encoder features with features from a frozen semantic segmentation encoder: $$F_{d}^{i} = \phi \big(PWConv_{1 \times 1}(DWConv_{3 \times 3}(F^{i}))\big), \tag{1}$$9 where $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$0 is the number of feature levels (Liu et al., 6 Nov 2025). The paper describes this as introducing semantic knowledge into the encoder to improve object recognition and boundary perception. The training description indicates that the semantic encoder is pretrained for semantic segmentation and then frozen. A plausible implication is that this branch functions as training-time guidance rather than as part of the deployed inference graph, although the inference-time use of the semantic branch is not explicitly specified (Liu et al., 6 Nov 2025).

4. Empirical results and ablation structure

The model is evaluated on NYUv2, KITTI, and iBims-1, with the last used for zero-shot generalization after training only on NYUv2 (Liu et al., 6 Nov 2025). The reported metrics are Abs Rel, RMSE, $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$1, $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$2, $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$3, and the depth boundary error accuracy $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$4, where lower boundary error is better.

On NYUv2, BoRe-Depth reports Abs Rel $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$5, RMSE $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$6, $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$7 $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$8, $$F_{ce}^{i} = F^{i} + Conv_{1 \times 1}(F_{d}^{i}), \tag{2}$$9 $DWConv_{3\times3}$0, $DWConv_{3\times3}$1 $DWConv_{3\times3}$2, and $DWConv_{3\times3}$3 $DWConv_{3\times3}$4, with 8.7M parameters (Liu et al., 6 Nov 2025). In the comparison table reproduced in the paper, these are better than DistDepth’s Abs Rel $DWConv_{3\times3}$5, RMSE $DWConv_{3\times3}$6, and boundary error $DWConv_{3\times3}$7, and better than GasMono’s Abs Rel $DWConv_{3\times3}$8, RMSE $DWConv_{3\times3}$9, and boundary error $PWConv_{1\times1}$0, while using far fewer parameters than either model.

On KITTI, BoRe-Depth reports Abs Rel $PWConv_{1\times1}$1, RMSE $PWConv_{1\times1}$2, $PWConv_{1\times1}$3 $PWConv_{1\times1}$4, $PWConv_{1\times1}$5 $PWConv_{1\times1}$6, $PWConv_{1\times1}$7 $PWConv_{1\times1}$8, and $PWConv_{1\times1}$9 $\phi$0 (Liu et al., 6 Nov 2025). In the provided comparison, these values outperform the listed lightweight models across all shown depth metrics and the boundary metric, including Lite-Mono, Dynamo-Depth, SC-DepthV3, and WeatherDepth.

The zero-shot iBims-1 evaluation is particularly important for the paper’s generalization claim. Trained only on NYUv2, BoRe-Depth reports Abs Rel $\phi$1, RMSE $\phi$2, $\phi$3 $\phi$4, $\phi$5 $\phi$6, $\phi$7 $\phi$8, and $\phi$9 $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$0 (Liu et al., 6 Nov 2025). These are stronger than the corresponding entries for DistDepth, GasMono, SC-DepthV3, and GAM-Depth in the paper’s comparison table, while still using the smallest parameter count among that group.

The ablation studies isolate the contributions of EFAF and the semantic training schedule. Starting from a 7.3M-parameter baseline, adding full EFAF raises the model size to 8.7M and improves NYUv2 performance from Abs Rel $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$1 to $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$2, RMSE from $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$3 to $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$4, $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$5 from $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$6 to $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$7, and $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$8 from $$F_{fusion}^{i} = \begin{cases} Conv(F_{ce}^{i} \oplus F_{fusion}^{i+1}), & i=1,2,3,\ Conv(F_{ce}^{i} \oplus F_{ce}^{i+1}), & i=4, \end{cases} \tag{3}$$9 to $\oplus$0 (Liu et al., 6 Nov 2025). Removing either the high-level or low-level SCE branch weakens these gains, which the paper uses to argue that both sides of adjacent-scale enhancement matter.

A second ablation compares semantic strategies. Relative to the EFAF baseline, adding a semantic decoder yields modest gains, while applying $\oplus$1 in stage 1 helps somewhat more. The best result comes from adding $\oplus$2 only in stage 2, improving Abs Rel from $\oplus$3 to $\oplus$4, RMSE from $\oplus$5 to $\oplus$6, $\oplus$7 from $\oplus$8 to $\oplus$9, and $P_t$0 from $P_t$1 to $P_t$2 (Liu et al., 6 Nov 2025). The paper interprets this as evidence that semantic knowledge is most useful after the model has already learned coarse geometric structure.

5. Embedded deployment and operational profile

Embedded deployment is a defining constraint of BoRe-Depth rather than a post hoc benchmark (Liu et al., 6 Nov 2025). The paper reports a total model size of 8.7M parameters and deployment on NVIDIA Jetson Orin at 50.7 FPS. This is presented as evidence that boundary-aware self-supervised monocular depth estimation can be made practical for real-time unmanned-system perception without resorting to large models.

The design choices supporting this deployment profile are explicit. The encoder is MPViT-tiny with 5.8M parameters, and the decoder uses depth-wise and point-wise convolutions to control computational cost. The overall architecture remains compact even after adding EFAF and the semantic-guided training strategy. The paper’s qualitative discussion further claims that BoRe-Depth produces clearer object contours, more accurate dense depth maps, and improved point cloud quality in comparison with prior lightweight models (Liu et al., 6 Nov 2025).

At the same time, several deployment details are not specified in the provided description: the exact Jetson Orin variant, input resolution for the speed benchmark, numerical precision, TensorRT usage, latency in milliseconds, and the software stack used for measurement are not given (Liu et al., 6 Nov 2025). This suggests that the headline embedded result is operationally meaningful, but that hardware-specific reproduction would require the accompanying code or additional implementation notes.

6. Terminological ambiguity, misconceptions, and limitations

A common misconception is to treat BoRe-Depth as only a decoder modification or only a boundary loss. In the paper, it is neither. The method is defined by the combination of a lightweight MPViT-based encoder–decoder, EFAF-based adjacent-scale fusion, pseudo-depth boundary supervision, and second-stage semantic feature alignment (Liu et al., 6 Nov 2025). Another misconception is to assume that semantic guidance is implemented as a standard multitask shared-encoder depth-and-segmentation architecture. The ablation results show that the paper instead favors a frozen semantic encoder guiding the depth encoder through feature similarity, with the strongest gains obtained when semantic loss is introduced only in the second stage (Liu et al., 6 Nov 2025).

The name itself also requires care. In the cited literature, “BoRe-Depth” is explicitly the title of the embedded monocular depth estimation model (Liu et al., 6 Nov 2025). By contrast, in “Depth Matching Method Based on ShapeDTW for Oil-Based Mud Imager,” the term does not appear as the paper’s algorithmic name and is described only as an external or project-level label for a ShapeDTW-based upper/lower pad depth alignment workflow (Li et al., 1 Dec 2025). The latter is a borehole-image depth-matching method rather than a monocular depth estimation architecture.

Several limitations remain explicit or implicit in the BoRe-Depth paper. Pseudo-depth labels are stated to possess limited accuracy and inherent model bias (Liu et al., 6 Nov 2025). The training description omits the exact form of $P_t$3, the exact similarity function in $P_t$4, input resolution, crop and augmentation policy, batch size, optimizer type, and several deployment details. The excerpt also does not provide a dedicated failure-case analysis for reflective surfaces, dynamic-scene violations, or severe low-texture regions. This suggests that the paper’s main contribution is an architectural and training formulation for the embedded self-supervised setting, rather than a full characterization of all robustness regimes.

In the present literature, BoRe-Depth is therefore best understood as a lightweight self-supervised monocular depth estimation system whose distinctive contribution is to elevate boundary quality to a primary design target while preserving embedded feasibility. Its technical identity lies in the joint use of EFAF, pseudo-depth boundary alignment, and second-stage semantic feature guidance, and its empirical profile is defined by low parameter count, real-time Jetson deployment, and consistently improved depth-boundary metrics across indoor, outdoor, and zero-shot evaluations (Liu et al., 6 Nov 2025).