Ordinal Depth Ranking Loss

mediumPRODeep Depth EstimationDeep LearningCV: Depth Estimation

Compute the ordinal (ranking) depth loss.

Ordinal depth loss enforces that the relative ordering of depths is preserved, even if absolute depths are wrong. For each pair of pixels $(i, j)$ with $i < j$ :

$L_{i,j} = \max(0, -\text{sign}(Z_i - Z_j) \cdot (\hat{Z}_i - \hat{Z}_j) + \tau)$

where:

$Z$ is ground truth depth
$\hat{Z}$ is predicted depth
$\tau$ is a margin
$\text{sign}()$ returns -1, 0, or 1

Pairs with equal ground-truth depth ( $Z_i = Z_j$ ) are skipped and contribute no loss. The total loss is the sum of $L_{i,j}$ over all remaining pairs, rounded to 4 decimal places.

The loss penalizes predictions that violate the ground truth ordering, or respect it by less than the margin τ.

Ordinal loss is useful when absolute depth is hard to obtain but relative depth (which object is closer) is known.

Example

Input

ordinal_loss([1, 2, 3], [1, 2, 3], 0.1)

Output

0.0

Reasoning

pred and gt hold one depth per pixel, and margin is $\tau = 0.1$ . Every unordered pair of pixels is compared once: $(0,1)$ , $(0,2)$ , $(1,2)$ .
Pair $(0,1)$ : ground truth gives $Z_0 - Z_1 = 1 - 2 = -1$ , so the sign is $-1$ (pixel 0 is closer). The predicted gap $\hat{Z}_0 - \hat{Z}_1 = -1$ agrees, so $L_{0,1} = \max(0, -(-1)(-1) + 0.1) = \max(0, -0.9) = 0$ .
Pair $(0,2)$ : sign $-1$ , predicted gap $-2$ , so $\max(0, -2 + 0.1) = 0$ . Pair $(1,2)$ : sign $-1$ , gap $-1$ , so $\max(0, -0.9) = 0$ .
The total is $0 + 0 + 0$ , printed as 0.0: every predicted gap points the right way and exceeds the margin $0.1$ .
Notice a pair only costs something when its predicted gap points the wrong way or is right by less than $\tau$ . Reversing the prediction to $[3, 2, 1]$ makes every pair violate: $1.1 + 2.1 + 1.1 = 4.3$ .
Pairs whose ground-truth depths are equal carry no ordering and add no loss.

Constraints

pred: list of predicted depths
gt: list of ground truth depths
margin: ranking margin (τ)
Return total ordinal loss rounded to 4 decimal places

Learn this in

Depth Estimation3D Computer Vision

More from CV: Depth Estimation

Plane Sweep Stereohard Block Matching Disparityhard Epipolar Line Visualizationhard Depth Map Normalizationeasy Scale-Invariant Depth Lossmedium

Concept primerPREVIEW

Ordinal Depth Ranking Loss: Background & Implementation Guide

Background Knowledge

Depth Estimation and Relative Ordering: In computer vision, monocular depth estimation predicts spatial depth from a single 2D image. While absolute depth values are difficult to predict accurately from monocular images, the relative ordering of depths—which pixels are closer or farther—is often more reliable and easier to learn. This is the core insight behind ordinal depth loss: instead of penalizing absolute depth errors, we penalize violations of the ground truth depth ordering between pixel pairs.

Ordinal Regression in Computer Vision: Ordinal regression treats depth prediction as a ranking problem rather than a pure regression problem. The key difference is that ordinal methods exploit the natural ordering structure in the data (closer < farther) to improve learning efficiency and generalization. This approach has proven effective in depth estimation because it aligns with how humans perceive depth—we naturally judge relative distances before estimating absolute values.

Why Margin-Based Ranking Loss?: The margin parameter $\tau$ allows the loss to tolerate small ordering violations without penalty. This is crucial because: (1) it prevents the network from over-fitting to noisy ground truth labels, (2) it provides a "safety zone" where predictions don't need to be perfect, and (3) it encourages the network to focus on clear ordering violations rather than minor depth fluctuations.

Algorithm/Approach

The ordinal depth ranking loss operates on pairwise comparisons of pixels:

Identify the ground truth ordering: For each pair of pixels $(i, j)$ , determine which is closer using $\text{sign}(Z_i - Z_j)$
Check predicted ordering: Compute the predicted depth difference $($ \hat{Z} $_i -$ \hat{Z} $_j)$
Measure violation: Multiply the sign of ground truth ordering by the predicted difference to detect violations
Apply margin: Only penalize violations that exceed the margin threshold $\tau$
Aggregate: Sum losses across all pixel pairs (or a sampled subset for efficiency)

This is a ranking-based loss function, similar to approaches used in object detection and metric learning.

Step-by-Step Implementation Strategy

Step 1: Understand the Loss Formula

Break down the formula component by component:

$\text{sign}(Z_i - Z_j)$ produces: $+1$ if $Z_i > Z_j$ , $-1$ if $Z_i < Z_j$ , $0$ if equal
$($ \hat{Z} $_i -$ \hat{Z} $_j)$ is the predicted depth difference
The product $\text{sign}(Z_i - Z_j) \cdot ($ \hat{Z} $_i -$ \hat{Z} $_j)$ is positive when predictions agree with ground truth, negative when they disagree
Negating this product: $-$ \text{sign} $(Z_i - Z_j) \cdot ($ \hat{Z} $_i -$ \hat{Z} $_j)$ becomes negative for correct orderings and positive for violations
$\max(0, \cdot + \tau)$ applies the margin: only penalize when violation exceeds $\tau$

PRO

Continue the full explanation

You're reading the free preview. Unlock the complete walkthrough, the code editor, test runner and reference solution with Premium.

PRO

Editor locked

The code editor is locked for Pro problems. It is only available for free problems. Please upgrade to gain access to the code editor for all problems.

solution.py

Python 3.13

Test results