notesum.ai

Published at December 4

Optimizing Dense Visual Predictions Through Multi-Task Coherence and Prioritization

cs.CV

cs.AI

Released Date: December 4, 2024

Authors: Maxime Fontana¹, Michael Spratling², Miaojing Shi³

Aff.: ¹Department of Informatics, King's College London; ²Department of Behavioural and Cognitive Sciences, University of Luxembourg; ³College of Electronic and Information Engineering, Tongji University

Arxiv: http://arxiv.org/pdf/2412.03179v1

Model	Semseg (mIoU) $\uparrow$	Depth (RMSE) $\downarrow$	Normal (mErr) $\downarrow$
Cross-Stitch [27]	36.34	0.6290	20.88
PAP [54]	36.72	0.6178	20.82
PSD [56]	36.69	0.6246	20.87
PAD-Net [45]	36.61	0.6270	20.85
MTI-Net [39]	45.97	0.5365	20.27
InvPT [50]	53.56	0.5183	19.04
DeMT [47]	51.50	0.5474	20.02
MLoRE [49]	55.96	0.5076	18.33
Bi-MTDP [33]	54.86	0.5150	19.50
STL ${}_{\text{Semseg}}$	53.20	-	-
STL ${}_{\text{Depth}}$	-	0.4923	-
STL ${}_{\text{Normal}}$	-	-	19.22
MT-CP	56.25	0.4316	18.60