💻 computer science

Sphere-Depth: A Benchmark for Depth Estimation Methods with Varying Spherical Camera Orientations

This paper introduces Sphere-Depth, a new benchmark designed to evaluate the robustness of monocular depth estimation models for equirectangular images against unintentional camera pose variations, revealing that even spherical-aware models suffer significant performance degradation when facing such perturbations.

Original authors: Soulayma Gazzeh, Giuseppe Mazzola, Liliana Lo Presti, Marco La Cascia

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Soulayma Gazzeh, Giuseppe Mazzola, Liliana Lo Presti, Marco La Cascia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are wearing a high-tech VR headset that gives you a full 360-degree view of the world. To make this world feel real, the headset needs to know exactly how far away everything is—the chair in front of you, the wall behind you, and the ceiling above you. This is called depth estimation.

Most of the "brains" (AI models) we currently use to do this are trained with one big assumption: that the camera is always perfectly level.

The Problem: The "Drunken Robot" Scenario

Think of a professional photographer. They use a tripod to keep their camera perfectly straight and level. They are in a "canonical pose"—the gold standard of stability. Most AI models are trained to be experts only when the camera is behaving like that professional photographer.

But now, imagine a robot walking through a bumpy construction site or a drone flying through a gust of wind. The camera isn't level anymore; it’s tilting left, right, up, and down. It’s like a photographer trying to take a perfect photo while riding a unicycle on a cobblestone street.

The researchers found that when the camera tilts (what they call pitch and roll), the AI gets "dizzy." Even the smartest AI models, which were specifically designed to understand 360-degree spherical images, suddenly start guessing the distances wrong. They might think a wall that is 5 meters away is actually 10 meters away, which could cause a robot to crash.

The Solution: "Sphere-Depth" (The Ultimate Obstacle Course)

To fix this, the researchers created a new benchmark called Sphere-Depth.

Think of Sphere-Depth as a stress test or an obstacle course for AI. Instead of just testing the AI in a calm, perfect laboratory, they intentionally "shake" the virtual camera. They tilt it at different angles to see which AI models stay focused and which ones lose their sense of direction.

They also created a clever way to grade the AI. Since different AI models "speak" in different scales (one might measure in inches while another uses centimeters), the researchers created a "Universal Translator" (a calibration protocol). This allows them to compare all the models fairly, ensuring they are all being judged on the same playing field.

The Results: Who Passed the Test?

The researchers put several "students" (AI models) through this obstacle course:

The Specialists (Spherical-aware models): These are the students who studied 360-degree geometry. Even though they were prepared, they still stumbled when the camera tilted too much. One student, ACDNet, was the star pupil—it was the most reliable even when things got shaky.
The Generalists (Perspective models): These are students trained on regular, flat photos (like the ones on your phone). When they tried to look at 360-degree views, they performed poorly, like someone trying to read a map that has been wrapped around a basketball.

Why does this matter?

If we want to build truly autonomous cars, delivery drones, or helpful home robots, we can't rely on "perfect" cameras. We need AI that is "pose-invariant"—meaning it can tell how far away a door is, even if the robot is wobbling, tilting, or bumping into things.

This paper provides the "training manual" and the "testing ground" to help engineers build AI that doesn't get motion sickness!

Technical Summary: Sphere-Depth Benchmark

Title: Sphere-Depth: A Benchmark for Depth Estimation Methods with Varying Spherical Camera Orientations
Authors: Soulayma Gazzeh, Giuseppe Mazzola, Liliana Lo Presti, and Marco La Cascia (University of Palermo)

1. Problem Statement

Monocular depth estimation from spherical (360°) images is vital for robotic navigation and immersive scene understanding. Most existing models—both perspective-based and spherical-aware—operate under the "canonical pose" assumption. This assumes the camera's equatorial plane is perfectly parallel to the floor (gravity-aligned).

In real-world robotics, however, mobile platforms (drones, autonomous vehicles, etc.) experience unintentional pitch and roll perturbations due to motion. These variations cause significant geometric distortions in Equirectangular Projections (ERP), which current models are not trained to handle, leading to unpredictable depth estimation errors.

2. Methodology

The authors propose a systematic framework to evaluate how sensitive depth estimation models are to these orientation changes.

The Sphere-Depth Benchmark: A novel dataset consisting of high-resolution indoor equirectangular images. The authors identified 37 landmarks with known 3D coordinates to serve as ground truth.
Simulated Perturbations: To test robustness, the researchers applied controlled rotations:
- Large Variations: Pitch and roll angles ranging from $-40^\circ$ to $+40^\circ$ .
- Small Variations: Subtle changes between $\pm 2^\circ$ to simulate minor sensor misalignments.
Depth Calibration Protocol: Since monocular models often output relative depth rather than metric depth, the authors introduced a learnable scaling factor ( $\lambda$ ). This factor aligns predicted depth maps with ground truth using a least-squares approach on a subset of landmarks, ensuring a fair comparison between models with different output scales.
Models Evaluated:
- Perspective-based: Depth Anything v2 (via cubemap projection).
- Spherical-aware: DepthAnywhere, BiFuse++, ACDNet, and SliceNet.

3. Key Contributions

A Pose-Aware Benchmark: The first benchmark specifically designed to measure the sensitivity of 360° depth models to pitch and roll variations.
Standardized Evaluation Protocol: A depth calibration method that uses a scaling factor to allow meaningful error comparison across diverse model architectures.
Comparative Sensitivity Analysis: A detailed study revealing that even models explicitly designed for spherical geometry suffer significant performance degradation when the camera is not in a canonical pose.

4. Experimental Results

Performance Degradation: All models showed increased error as pose variation increased. ACDNet performed best on gravity-aligned images ( $\epsilon = 0.21$ ), but even it saw error increases under deformation.
Model Specifics:
- SliceNet showed unusual, asymmetric error patterns and struggled even with small pose changes.
- DepthAnywhere, BiFuse++, and SliceNet struggled significantly with high-deformation scenarios.
- Perspective Models (Depth Anything): When used via cubemap projection, these models failed to account for the scene's global geometry, resulting in high errors ( $\epsilon = 1.72$ ) due to inaccuracies at the cube face borders.
Depth Range Bias: Analysis via LOWESS curves showed that models have different "sweet spots." For example, SliceNet is effective for close-range depth ( $<2$ m), while DepthAnywhere is more stable for long-range depth ( $>3$ m).
Sensitivity: Even "small" perturbations ( $\pm 2^\circ$ ) caused a measurable performance drop (9–19% error increase) in top-performing models like ACDNet.

5. Significance

This paper exposes a critical gap in current 360° computer vision research: the lack of pose-invariance. By providing a reproducible benchmark and a new evaluation protocol, the authors provide a roadmap for the development of more robust, geometry-invariant depth estimation algorithms that can function reliably on moving robotic platforms in unconstrained environments.