Depth-Pushing Loss: A Cross-Domain Perspective

• Depth-Pushing Loss is a family of training objectives that impose directional bias on depth estimates to steer predictions away from geometrically invalid or ambiguous states.
• It is applied across domains such as self-supervised depth estimation, depth completion, multi-focus fusion, rendering, and statistical depth by tailoring losses like negative depth, asymmetric penalties, and correlation losses.
• Empirical results on benchmarks like KITTI demonstrate that integrating depth-pushing mechanisms leads to improved accuracy, sharper depth boundaries, and more robust handling of occlusions.

Searching arXiv for the cited papers to ground the article in current records. arXiv search: "Improved Point Transformation Methods For Self-Supervised Depth Prediction" Depth-Pushing Loss denotes a family of training objectives that impose explicit directional pressure on estimates involving depth, depth ordering, depth-of-field, or depth-conditioned supervision. In the literature surveyed here, the expression is not a single standardized loss with a single formula; rather, it describes a recurring design pattern in which optimization is biased away from geometrically invalid, structurally ambiguous, defocused, or under-supervised depth states. Representative instances include the negative depth loss for self-supervised monocular depth prediction, asymmetric foreground/background losses for depth completion at occlusion boundaries, decision-level focus-property constraints for multi-focus photoacoustic microscopy, multi-scale correlation losses for sparse-view 3D Gaussian Splatting, depth-weighted supervision for object detection, and risk-based “loss depths” in statistical learning (Ziwen et al., 2021, Imran et al., 2021, Zhou et al., 29 May 2025, Lu et al., 28 May 2025, Sbeyti et al., 5 Feb 2026, Castellanos et al., 11 Jul 2025).

1. Scope and recurring structure

A common feature of depth-pushing objectives is that they do not merely regularize outputs toward smoothness or metric agreement. Instead, they encode a preferred direction in the error landscape. In geometric settings, that direction may be “toward positive transformed depth” or “toward the correct side of an occlusion boundary.” In imaging settings, it may be “toward all-in-focus fusion across depth.” In detection, it may be “toward greater learning pressure on distant objects.” In statistical depth, it becomes “toward greater centrality with respect to a distribution.”

Paper	Domain	Depth-pushing mechanism
(Ziwen et al., 2021)	Self-supervised monocular depth prediction	Negative depth loss on in-frame, negative-depth projections
(Imran et al., 2021)	Depth completion	ALE/RALE asymmetric losses for twin-surface extrapolation
(Zhou et al., 29 May 2025)	OR-PAM multi-focus fusion	Decision-level focus property perceptual loss
(Lu et al., 28 May 2025)	Sparse-view 3DGS	Cascade Pearson Correlation Loss
(Sbeyti et al., 5 Feb 2026)	Object detection	Depth-Based Loss Weighting and Loss Stratification
(Castellanos et al., 11 Jul 2025)	Statistical data depth	Loss depths as minimum classification risk

This diversity is consequential. A frequent misconception is to treat Depth-Pushing Loss as a canonical module analogous to a standard photometric loss or a standard contrastive loss. The surveyed work suggests the opposite: the term is best understood as an umbrella description for objectives that encode a depth-aware optimization bias, with substantially different semantics across subfields.

2. Geometric barrier losses in self-supervised depth prediction

In self-supervised monocular depth estimation from stereo or egomotion pairs, the depth-pushing mechanism is explicit. The method in "Improved Point Transformation Methods For Self-Supervised Depth Prediction" defines a target-frame pixel $u=(u,v)$ with homogeneous coordinates $\tilde u=[u,v,1]^T$, intrinsics $K$, predicted depth $d_t(u)$, and relative pose $T_{t\rightarrow s}=(R,t)$. The transformation pipeline is

$$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$

followed by projection

$$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$

Negative transformed depth corresponds to points that end up behind the source camera. The paper operationalizes the in-frame negative-depth set as

$$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$

and defines the negative depth loss

$L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$

Because $d_{i,j}<0$ on $\tilde u=[u,v,1]^T$0, this is equivalent to summing $\tilde u=[u,v,1]^T$1. The gradient is therefore constant and negative with respect to any negative transformed depth, pushing $\tilde u=[u,v,1]^T$2 upward toward positive values. The mechanism is barrier-like but implemented without a margin: the condition is exactly $\tilde u=[u,v,1]^T$3. This is crucial early in training, when overly shallow predictions may transform behind the second camera; if these samples are only masked out, photometric supervision collapses over a large fraction of pixels and optimization can stagnate.

The loss is integrated into the full self-supervised objective

$\tilde u=[u,v,1]^T$4

with $\tilde u=[u,v,1]^T$5, $\tilde u=[u,v,1]^T$6, $\tilde u=[u,v,1]^T$7, and $\tilde u=[u,v,1]^T$8 in the KITTI experiments. The training protocol excludes points in $\tilde u=[u,v,1]^T$9 from $K$0, $K$1, and $K$2; those pixels contribute only to $K$3. Out-of-frame projections are excluded from all losses, and a differentiable z-buffer is then applied to the remaining in-frame, positive-depth points to resolve visibility exactly. This makes the z-buffer and $K$4 complementary: the z-buffer handles occlusions among valid points, while $K$5 converts in-frame, behind-camera projections into valid candidates for photometric supervision.

The empirical effect is measurable on KITTI (Eigen split). The baseline reports Abs Rel $K$6, Sq Rel $K$7, RMSE $K$8, RMSE log $K$9, $d_t(u)$0 $d_t(u)$1, $d_t(u)$2 $d_t(u)$3, and $d_t(u)$4 $d_t(u)$5. Adding the negative depth loss without occlusion handling yields Abs Rel $d_t(u)$6, Sq Rel $d_t(u)$7, RMSE $d_t(u)$8, RMSE log $d_t(u)$9, $T_{t\rightarrow s}=(R,t)$0 $T_{t\rightarrow s}=(R,t)$1, $T_{t\rightarrow s}=(R,t)$2 $T_{t\rightarrow s}=(R,t)$3, and $T_{t\rightarrow s}=(R,t)$4 $T_{t\rightarrow s}=(R,t)$5, improving all metrics except a negligible change in Abs Rel. With z-buffer insertion at epoch $T_{t\rightarrow s}=(R,t)$6 of $T_{t\rightarrow s}=(R,t)$7 plus the negative depth loss, the best row reports Abs Rel $T_{t\rightarrow s}=(R,t)$8, Sq Rel $T_{t\rightarrow s}=(R,t)$9, RMSE $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$0, RMSE log $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$1, $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$2 $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$3, $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$4 $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$5, and $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$6 $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$7 (Ziwen et al., 2021).

3. Asymmetric depth pushing at occlusion boundaries

In depth completion, the failure mode is not behind-camera reprojection but depth smearing across occlusion boundaries. "Depth Completion with Twin Surface Extrapolation at Occlusion Boundaries" addresses this by replacing single-surface interpolation with a twin-surface representation. The network predicts foreground depth $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$8, background depth $$X_t(u)=d_t(u)K^{-1}\tilde u,\quad X_s(u)=RX_t(u)+t,\quad z_s(u)=e_3^T X_s(u),$$9, and a foreground selection weight $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$0, with the final fused depth

$$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$1

The key depth-pushing mechanism is the use of asymmetric losses on the two surfaces. With prediction errors $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$2 and $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$3, and asymmetry parameter $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$4, the paper defines

$$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$5

$$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$6

Foreground supervision uses $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$7; background supervision uses $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$8. This creates a directional bias. For $$\tilde u_s(u)=\pi(KX_s(u)),\qquad \pi([x,y,z]^T)=[x/z,y/z]^T.$$9, overestimating depth is penalized with slope $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$0 and underestimating with slope $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$1, so the foreground branch is pushed toward shallower solutions. For $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$2, underestimating depth is penalized strongly and overestimating weakly, so the background branch is pushed toward deeper solutions.

The paper further characterizes the ambiguity analytically. For a pixel with two possible true depths $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$3 and probabilities $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$4 and $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$5, the expected ALE is minimized at the foreground depth $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$6 if

$$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$7

while the expected RALE is minimized at the background depth $$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$8 if

$$\mathcal N \triangleq \{(i,j)\mid d_{i,j}<0 \wedge 0\le i<w \wedge 0\le j<h\},$$9

This establishes a precise sense in which asymmetry implements a depth-pushing estimator: it biases ambiguous predictions toward opposite sides of a step discontinuity rather than toward the average.

Supervision is multi-scale: $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$0 with

$L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$1

The KITTI schedule uses three phases over $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$2 epochs: epochs $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$3–$L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$4, $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$5; epochs $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$6–$L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$7, $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$8, $L_{nd}=\sum_{(i,j)\in \mathcal N}|d_{i,j}|.$9; epochs $d_{i,j}<0$0–$d_{i,j}<0$1, $d_{i,j}<0$2, $d_{i,j}<0$3. The reported good compromise is $d_{i,j}<0$4.

Empirically, the method improves boundary-sensitive error measures. On KITTI test/validation, TWISE reports MAE $d_{i,j}<0$5, iMAE $d_{i,j}<0$6, RMSE $d_{i,j}<0$7, and iRMSE $d_{i,j}<0$8. An ablation on the MultiStack backbone shows TWISE with MAE $d_{i,j}<0$9, RMSE $\tilde u=[u,v,1]^T$00, TMAE $\tilde u=[u,v,1]^T$01, and TRMSE $\tilde u=[u,v,1]^T$02, compared with L1 at MAE $\tilde u=[u,v,1]^T$03, RMSE $\tilde u=[u,v,1]^T$04, TMAE $\tilde u=[u,v,1]^T$05, and TRMSE $\tilde u=[u,v,1]^T$06. The learned fusion variable is also essential: using only $\tilde u=[u,v,1]^T$07 gives MAE $\tilde u=[u,v,1]^T$08, only $\tilde u=[u,v,1]^T$09 gives $\tilde u=[u,v,1]^T$10, simple averaging gives $\tilde u=[u,v,1]^T$11, learned $\tilde u=[u,v,1]^T$12 without color gives $\tilde u=[u,v,1]^T$13, and learned $\tilde u=[u,v,1]^T$14 with color gives the best MAE $\tilde u=[u,v,1]^T$15 (Imran et al., 2021).

4. Depth-of-field extension as focus-property pushing

In optical-resolution photoacoustic microscopy, the depth-pushing idea shifts from metric depth prediction to depth-of-field extension. "Dc-EEMF: Pushing depth-of-field limit of photoacoustic microscopy via decision-level constrained learning" frames multi-focus fusion as the construction of an all-in-focus image from two source images $\tilde u=[u,v,1]^T$16 and $\tilde u=[u,v,1]^T$17 acquired at different focal planes. The method, Dc-EEMF, is a lightweight Siamese CNN with feature extraction, artifact-resistant channel-wise spatial frequency fusion, and feature reconstruction. The depth-pushing component is the decision-level focus property perceptual loss $\tilde u=[u,v,1]^T$18, which compares the fused image’s focus properties against ground-truth focus property maps using a dual-input U-Net.

The total objective is

$\tilde u=[u,v,1]^T$19

with $\tilde u=[u,v,1]^T$20, $\tilde u=[u,v,1]^T$21, and $\tilde u=[u,v,1]^T$22. The perceptual term is

$\tilde u=[u,v,1]^T$23

with VGG19 features from layers $\tilde u=[u,v,1]^T$24. The structural term is

$\tilde u=[u,v,1]^T$25

and the focal frequency term is

$\tilde u=[u,v,1]^T$26

The role of $\tilde u=[u,v,1]^T$27 is decision-level rather than pixel-level: for each source $\tilde u=[u,v,1]^T$28, the pair $\tilde u=[u,v,1]^T$29 is fed to a dual-input U-Net, and an MSE is computed between the predicted focus property map and the corresponding ground-truth focus property map. This encourages the fused image to inherit the correct in-focus regions from each source, thereby pushing depth-of-field.

The feature fusion rule uses channel-wise spatial frequency, aggregated over an $\tilde u=[u,v,1]^T$30 window to resist artifact contamination near boundaries. The binary decision tensor is

$\tilde u=[u,v,1]^T$31

and the fused feature map is

$\tilde u=[u,v,1]^T$32

This rule favors sharper channel-wise features while suppressing misalignment artifacts.

The reported quantitative and practical characteristics are unusually explicit. The U-Net for focus-property detection reaches mean IoU $\tilde u=[u,v,1]^T$33, with minor misclassifications near boundaries. Dc-EEMF is trained end-to-end in PyTorch 1.11.0 using Adam with $\tilde u=[u,v,1]^T$34, $\tilde u=[u,v,1]^T$35, $\tilde u=[u,v,1]^T$36, learning rate $\tilde u=[u,v,1]^T$37 with $\tilde u=[u,v,1]^T$38 decay every $\tilde u=[u,v,1]^T$39 epochs, batch size $\tilde u=[u,v,1]^T$40, and $\tilde u=[u,v,1]^T$41 epochs, without post-processing. The model is lightweight at approximately $\tilde u=[u,v,1]^T$42 million parameters and processes $\tilde u=[u,v,1]^T$43 inputs in approximately $\tilde u=[u,v,1]^T$44 ms on an NVIDIA 2080Ti. On in vivo mouse brain data, fusing $\tilde u=[u,v,1]^T$45 at $\tilde u=[u,v,1]^T$46 and $\tilde u=[u,v,1]^T$47 at $\tilde u=[u,v,1]^T$48 computationally pushes DoF to approximately $\tilde u=[u,v,1]^T$49 while preserving acceptable transverse resolution; junction density and vessel density both increase significantly, with Mann-Whitney U Test values $\tilde u=[u,v,1]^T$50 and $\tilde u=[u,v,1]^T$51, respectively (Zhou et al., 29 May 2025).

5. Multi-scale geometry pushing in rendering and distant-object supervision

In sparse-view novel view synthesis, the depth-pushing objective can act directly on rendered geometry. "Learning Fine-Grained Geometry for Sparse-View Splatting via Cascade Depth Loss" introduces Hierarchical Depth-Guided Splatting (HDGS) and its Cascade Pearson Correlation Loss (CPCL). Let $\tilde u=[u,v,1]^T$52 be rendered depth from the current 3DGS model and $\tilde u=[u,v,1]^T$53 the monocular depth from DPT. A depth pyramid is constructed by average pooling, and each level is partitioned into non-overlapping patches. For patch $\tilde u=[u,v,1]^T$54 at level $\tilde u=[u,v,1]^T$55, centered vectors $\tilde u=[u,v,1]^T$56 and $\tilde u=[u,v,1]^T$57 are formed, and the patch correlation is

$\tilde u=[u,v,1]^T$58

with $\tilde u=[u,v,1]^T$59 and $\tilde u=[u,v,1]^T$60. Per-level correlation is $\tilde u=[u,v,1]^T$61, and the loss is

$\tilde u=[u,v,1]^T$62

with uniform weights $\tilde u=[u,v,1]^T$63. Because Pearson correlation is scale- and shift-invariant, CPCL aligns depth structure without requiring metric scale agreement between $\tilde u=[u,v,1]^T$64 and $\tilde u=[u,v,1]^T$65. Through the rendered depth

$\tilde u=[u,v,1]^T$66

its gradients act on splat positions, covariances, and opacities, pushing geometry toward monocular depth structure rather than raw absolute depth values. HDGS integrates CPCL with normalized L2 terms in local and global modes, using $\tilde u=[u,v,1]^T$67, $\tilde u=[u,v,1]^T$68, $\tilde u=[u,v,1]^T$69, $\tilde u=[u,v,1]^T$70, and $\tilde u=[u,v,1]^T$71. Under a $\tilde u=[u,v,1]^T$72-view sparse protocol, the reported results are PSNR $\tilde u=[u,v,1]^T$73, SSIM $\tilde u=[u,v,1]^T$74, LPIPS $\tilde u=[u,v,1]^T$75 on LLFF and PSNR $\tilde u=[u,v,1]^T$76, SSIM $\tilde u=[u,v,1]^T$77, LPIPS $\tilde u=[u,v,1]^T$78 on DTU; the best patch-scale configuration is $\tilde u=[u,v,1]^T$79, while adding $\tilde u=[u,v,1]^T$80 slightly hurts performance (Lu et al., 28 May 2025).

A different but related mechanism appears in object detection. "Depth as Prior Knowledge for Object Detection" does not name a Depth-Pushing Loss explicitly; the closest components are Depth-Based Loss Weighting (DLW) and Depth-Based Loss Stratification (DLS). DLW defines a monotone depth-dependent weight

$\tilde u=[u,v,1]^T$81

with recommended $\tilde u=[u,v,1]^T$82, and uses it in

$\tilde u=[u,v,1]^T$83

leaving negatives unweighted for stability. DLS instead splits supervision into close and distant strata using a split factor $\tilde u=[u,v,1]^T$84, and forms

$\tilde u=[u,v,1]^T$85

The defaults are $\tilde u=[u,v,1]^T$86, $\tilde u=[u,v,1]^T$87, and $\tilde u=[u,v,1]^T$88; for KITTI, the reported setting is $\tilde u=[u,v,1]^T$89 and $\tilde u=[u,v,1]^T$90. The inference-stage complement is Depth-Aware Confidence Thresholding (DCT), which adapts decision thresholds as a function of depth. This framework is motivated by a heteroscedastic model in which $\tilde u=[u,v,1]^T$91 grows quadratically with distance, biasing uniform SGD toward nearby objects. On KITTI with YOLOv11, baseline small-object mAP$\tilde u=[u,v,1]^T$92 $\tilde u=[u,v,1]^T$93 rises to $\tilde u=[u,v,1]^T$94 with DLW and $\tilde u=[u,v,1]^T$95 with DLS, while mAR$\tilde u=[u,v,1]^T$96 improves from $\tilde u=[u,v,1]^T$97 to $\tilde u=[u,v,1]^T$98 and $\tilde u=[u,v,1]^T$99. Across four benchmarks and two detectors, the abstract reports gains of up to $K$00 mAP$K$01 and $K$02 mAR$K$03, with inference recovery rates as high as $K$04 true versus false detections (Sbeyti et al., 5 Feb 2026).

6. Risk-based loss depths, limitations, and terminological heterogeneity

The broadest formalization appears in "Data Depth as a Risk," where depth no longer denotes scene geometry but statistical centrality. For a point $K$05 and distribution $K$06, the paper recalls the halfspace depth

$K$07

and shows that it can be written as a minimum classification risk under an artificial labeling

$K$08

For linear classifiers and $K$09–$K$10 loss, the theorem states

$K$11

This leads to a family of loss depths

$K$12

with empirical counterpart

$K$13

From this perspective, a depth-pushing loss is simply the negative of a regularized loss depth, so maximizing “depth” means making a target point harder to separate from the data distribution. The paper gives $K$14 convergence rates for logistic-regression depth and regularized kernel SVM depth, and emphasizes that classifier simplicity is essential: if $K$15 is too expressive, depth can collapse toward zero (Castellanos et al., 11 Jul 2025).

This statistical formulation clarifies an important terminological issue. In several of the vision papers discussed above, the label “Depth-Pushing Loss” is interpretive rather than author-specified: the OR-PAM paper states that it does not name a specific “Depth-Pushing Loss,” and the object-detection paper likewise states that it does not name one explicitly, though it identifies DLW and DLS as the closest mechanisms. A plausible implication is that the term is presently best treated as a cross-domain analytical category rather than a fixed named primitive (Zhou et al., 29 May 2025, Sbeyti et al., 5 Feb 2026).

The limitations reported across the literature are correspondingly task-specific. In self-supervised depth prediction, severe pose errors could temporarily create true negative-depth projections for otherwise valid geometry, though the loss is applied only to in-frame projections and is argued to be appropriate for KITTI stereo and egomotion (Ziwen et al., 2021). In twin-surface depth completion, KITTI ground truth is noisy near boundaries, so RMSE may worsen even when geometry is sharper, and fusion quality drops without RGB cues (Imran et al., 2021). In OR-PAM, residual focus-map misclassifications remain near boundaries despite MIoU $K$16 (Zhou et al., 29 May 2025). In sparse-view splatting, large local monocular-depth mistakes can push geometry in the wrong direction (Lu et al., 28 May 2025). In detection, overemphasizing distant objects can trade off against large-object performance, and DCT can inflate extra detections if validation-derived thresholds do not generalize (Sbeyti et al., 5 Feb 2026). In loss-depth theory, per-point optimization cost and classifier-class selection remain central practical concerns (Castellanos et al., 11 Jul 2025).

Taken together, these works support a precise but non-unified understanding: a Depth-Pushing Loss is any objective that introduces a depth-aware directional bias into optimization, whether by penalizing behind-camera projections, separating foreground and background hypotheses, enforcing focus-consistent fusion across depth, aligning rendered and monocular depth structure, upweighting distant supervision, or maximizing statistical centrality under a loss-defined depth notion.