English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Pre-training has been an important ingredient in developing strong monoculardepth estimation models in recent years. For instance, self-supervised learning(SSL) is particularly effective by alleviating the need for large datasets withdense ground-truth depth maps. However, despite these improvements, our studyreveals that the later layers of the SOTA SSL method are actually suboptimal.By examining the layer-wise representations, we demonstrate significant changesin these later layers during fine-tuning, indicating the ineffectiveness oftheir pre-trained features for depth estimation. To address these limitations,we propose MeSa, a comprehensive framework that leverages the complementarystrengths of masked, geometric, and supervised pre-training. Hence, MeSabenefits from not only general-purpose representations learnt via masked pretraining but also specialized depth-specific features acquired via geometricand supervised pre-training. Our CKA layer-wise analysis confirms that ourpre-training strategy indeed produces improved representations for the laterlayers, overcoming the drawbacks of the SOTA SSL method. Furthermore, viaexperiments on the NYUv2 and IBims-1 datasets, we demonstrate that theseenhanced representations translate to performance improvements in both thein-distribution and out-of-distribution settings. We also investigate theinfluence of the pre-training dataset and demonstrate the efficacy ofpre-training on LSUN, which yields significantly better pre-trainedrepresentations. Overall, our approach surpasses the masked pre-training SSLmethod by a substantial margin of 17.1% on the RMSE. Moreover, even withoututilizing any recently proposed techniques, MeSa also outperforms the mostrecent methods and establishes a new state-of-the-art for monocular depthestimation on the challenging NYUv2 dataset.

MeSa: Masked, Geometric, and Supervised Pre-training for Monocular Depth Estimation