Abstract
Self-supervised monocular depth estimation has seen significant progress in recent years, especially in outdoor environments. However, depth prediction results are not satisfying in indoor scenes where most of the existing data are captured with hand-held devices. As compared to outdoor environments, estimating depth of monocular videos for indoor environments, using self-supervised methods, results in two additional challenges: (i) the depth range of indoor video sequences varies a lot across different frames, making it difficult for the depth network to induce consistent depth cues for training; (ii) the indoor sequences recorded with handheld devices often contain much more rotational motions, which cause difficulties for the pose network to predict accurate relative camera poses. In this work, we propose a novel framework-MonoIndoor++ by giving special considerations to those challenges and consolidating a set of good practices for improving the performance of self-supervised monocular depth estimation for indoor environments. First, a depth factorization module with transformer-based scale regression network is proposed to estimate a global depth scale factor explicitly, and the predicted scale factor can indicate the maximum depth values. Second, rather than using a single-stage pose estimation strategy as in previous methods, we propose to utilize a residual pose estimation module to estimate relative camera poses across consecutive frames iteratively. Third, to incorporate extensive coordinates guidance for our residual pose estimation module, we propose to perform coordinate convolutional encoding directly over the inputs to pose networks. The proposed method is validated on a variety of benchmark indoor datasets, i.e., EuRoC MAV, NYUv2, ScanNet and 7-Scenes, demonstrating the state-of-the-art performance.
Introduction
While self-supervised monocular depth estimation has flourished in outdoor settings like autonomous driving, indoor scenes remain a significant challenge. The performance gap is primarily driven by two factors: Scale Volatility and Complex Ego-Motion. Unlike outdoor scenes where depth ranges are relatively consistent (e.g., the sky is always at infinity), indoor environments vary drastically—from cramped rooms to wide lobbies. Furthermore, indoor data is often captured via handheld devices or drones, resulting in erratic, high-rotation movements that standard pose estimation networks struggle to resolve.
Methodology
MonoIndoor++ addresses these hurdles through three core architectural innovations integrated into the self-supervised pipeline.
Depth Factorization with Transformers
Instead of predicting a single depth map, the system factorizes depth into a relative depth map and a global scale factor. A dedicated Transformer-based scale regression network predicts this global scale. The Transformer's attention mechanism allows the network to focus on specific informative regions, such as far corners or open doorways, to accurately gauge the scene's overall scale.
Residual Pose Estimation
Rather than estimating camera pose in a single step, MonoIndoor++ employs an iterative approach. An initial pose is predicted, the source image is warped to a "virtual view," and then a Residual PoseNet calculates the remaining error. This iterative refinement significantly improves robustness against the heavy rotations typical of handheld indoor footage.
Coordinate Convolutional Encoding
To provide the pose network with better geometric intuition, the authors utilize CoordConv. By explicitly concatenating $x$ and $y$ pixel coordinates as extra input channels, the network receives spatial guidance that helps it distinguish between translational and rotational pixel shifts, which is crucial in texture-less indoor environments.
Experimental Results
The model was evaluated on four diverse indoor benchmarks: NYUv2, EuRoC MAV, ScanNet, and 7-Scenes.
Quantitative Performance
On the NYUv2 dataset, MonoIndoor++ achieved an Absolute Relative Error (AbsRel) of 0.132, a substantial improvement over the Monodepth2 baseline (0.160). In terms of pose accuracy on the EuRoC MAV dataset, the residual pose module reduced rotational errors by more than half in several sequences, demonstrating the effectiveness of the iterative refinement strategy.
Qualitative Improvements
Visual results highlight the model's ability to produce sharper object boundaries and better reconstruct complex structures, such as chair legs or small furniture details, which are often blurred by competing models.
Generalization
A model trained on NYUv2 demonstrated impressive "zero-shot" performance on ScanNet, outperforming other state-of-the-art methods without requiring ScanNet-specific training data. This suggests that the learned depth cues and scale factorization are robust across different indoor environments.
Conclusion
MonoIndoor++ significantly advances self-supervised monocular depth estimation for indoor scenes by explicitly addressing scale volatility and complex camera motion. By replacing the missing global cues of outdoor scenes with Transformer-based scale prediction and refining pose estimation iteratively, the framework bridges the gap toward fully supervised performance. Future work may involve integrating semantic information or more advanced Transformer architectures to further enhance robustness in dynamic environments.
Summary updated