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

Pseudo-labels are widely employed in weakly supervised 3D segmentation taskswhere only sparse ground-truth labels are available for learning. Existingmethods often rely on empirical label selection strategies, such as confidencethresholding, to generate beneficial pseudo-labels for model training. Thisapproach may, however, hinder the comprehensive exploitation of unlabeled datapoints. We hypothesize that this selective usage arises from the noise inpseudo-labels generated on unlabeled data. The noise in pseudo-labels mayresult in significant discrepancies between pseudo-labels and modelpredictions, thus confusing and affecting the model training greatly. Toaddress this issue, we propose a novel learning strategy to regularize thegenerated pseudo-labels and effectively narrow the gaps between pseudo-labelsand model predictions. More specifically, our method introduces an EntropyRegularization loss and a Distribution Alignment loss for weakly supervisedlearning in 3D segmentation tasks, resulting in an ERDA learning strategy.Interestingly, by using KL distance to formulate the distribution alignmentloss, it reduces to a deceptively simple cross-entropy-based loss whichoptimizes both the pseudo-label generation network and the 3D segmentationnetwork simultaneously. Despite the simplicity, our method promisingly improvesthe performance. We validate the effectiveness through extensive experiments onvarious baselines and large-scale datasets. Results show that ERDA effectivelyenables the effective usage of all unlabeled data points for learning andachieves state-of-the-art performance under different settings. Remarkably, ourmethod can outperform fully-supervised baselines using only 1% of trueannotations. Code and model will be made publicly available athttps://github.com/LiyaoTang/ERDA.

All Points Matter: Entropy-Regularized Distribution Alignment for Weakly-supervised 3D Segmentation