notesum.ai

Published at December 10

RADIO Amplified: Improved Baselines for Agglomerative Vision Foundation Models

cs.CV

cs.AI

Released Date: December 10, 2024

Authors: Greg Heinrich¹, Mike Ranzinger¹, Hongxu, Yin, Yao Lu¹, Jan Kautz¹, Andrew Tao¹, Bryan Catanzaro¹, Pavlo Molchanov¹

Aff.: ¹NVIDIA

Arxiv: http://arxiv.org/pdf/2412.07679v1

Refer to caption

Layers	Aggre-	Head	ADE20k	Depth	Surf	Overall
	gation				Normals
31	N/A	Linear	52.47	82.9	57.0	61.215
7-15-23-31	Sparse	Linear	52.99	82.5	59.6	62.03
(0-9)-(10-19)-(20-30)-31	Dense	Linear	52.96	82.7	59.5	62.03
15-31	Sparse	Linear	52.90	83.1	59.6	62.12
(0-15)-(16-30)-31	Dense	DPT	54.27	85.4	60.7	63.65
15-31	Sparse	DPT	54.58	84.6	61.0	63.70
7-15-23-31	Sparse	DPT	55.19	85.9	61.6	64.46
(0-9)-(10-19)-(20-30)-31	Dense	DPT	54.28	85.5	62.3	64.08
3-7-11-15-19-23-27-31	Sparse	DPT	54.42	86.7	62.8	64.58
		TextVQA	ChartQA	DocVQA	InfoVQA	OCRBench
Last	N/A	63.6	23.4	47.0	33.8	42.0
7-15-23-31	Sparse	63.2	24.1	47.2	34.3	40.3
(0-9)-(10-19)-(20-30)-31	Dense	63.5	23.1	47.0	33.5	40.2