MaSS makes self-supervised monocular depth scale across mixed scenes
Original title: Learning to Discriminate Scene Structures Makes Self-Supervised Depth Learning Scalable
A decoupled structure pathway lets mixed-domain training help, with zero-shot tests on eight more datasets.
The paper argues self-supervised monocular depth models encode dataset-specific structural biases and conflict when trained on diverse scenes. MaSS adds a structural representation pathway, separate from contextual features, trained with a scene-structure discrimination objective. Trained and evaluated on four datasets, alone and jointly, it beat prior self-supervised methods and was the first in the study to gain consistently from mixed-domain training. Zero-shot results on eight further datasets, including out-of-distribution scenes, are reported, and code is public.
Unconfirmed: Dataset names and quantitative margins versus prior methods are not stated.
- Models were trained and evaluated on four datasets, individually and jointly.
- Zero-shot evaluation used eight additional datasets.
- The structure pathway is decoupled from contextual features.
- Authors say MaSS is first to consistently benefit from mixed-domain training.
- Source code is at https://mias.group/MaSS.
From the original article
1 min read at huggingface.co
Self-supervised monocular depth estimation is appealing for its potential to scale with unlabeled data, yet existing methods often degrade when trained on diverse scenes. This study finds that, under self-supervision, models tend to encode dataset-specific structural biases rather than transferable scene geometry, leading to cross-domain optimization conflicts.
By Mengtan Zhang, Yuchuan Cui, Yu Ma · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.