notesum.ai

Published at December 9

Elastic-DETR: Making Image Resolution Learnable with Content-Specific Network Prediction

cs.CV

cs.AI

Released Date: December 9, 2024

Authors: Daeun Seo¹, Hoeseok Yang², Sihyeong Park³, Hyungshin Kim¹

Aff.: ¹Chungnam National University; ²Santa Clara University; ³Korea Electronics Technology Institute

Arxiv: http://arxiv.org/pdf/2412.06341v1

Refer to caption

	Strategy	AP	AP₅₀	AP₇₅	AP_s	AP_m	AP_l	GFLOPs	# Params	# Epochs
DETR-R50 [3]	MS training	42.0	62.4	44.2	20.5	45.8	61.1	86	41M	500
Conditional-DETR-R50 [23]	MS training	43.0	64.0	45.7	22.7	46.7	61.5	90	44M	108
Anchor DETR-R50 [40]	MS training	42.1	63.1	44.9	22.3	46.2	60.0	-	39M	50
DAB-DETR-R50 [21]	MS training	42.2	63.1	44.7	21.5	45.7	60.3	94	44M	50
DN-DETR-R50 [15]	MS training	44.1	64.4	46.7	22.9	48.0	63.4	94	44M	50
DN-DETR-R50^∗	MS training	44.6	64.8	47.7	23.9	48.5	63.3	-	-	50
Elastic-DETR-S	Ours ( $\tau_{max}=1.25$ )	43.6	63.9	46.3	21.8	47.3	64.1	70	45M	50
Elastic-DETR-M	Ours ( $\tau_{max}=1.50$ )	45.1	65.4	48.2	23.6	49.0	64.1	99	45M	50
Elastic-DETR-B	Ours ( $\tau_{max}=1.75$ )	46.1	66.5	49.6	26.5	49.6	64.0	134	45M	50
Elastic-DETR-L^†	Ours ( $\tau_{max}=2.00$ )	47.0	67.0	50.8	27.2	51.4	65.0	170	45M	50
Elastic-DETR-H^†	Ours ( $\tau_{max}=2.25$ )	47.6	67.4	51.2	28.1	51.3	64.6	209	45M	50
DETR-Swin-T [30]	MS training	34.1	55.1	35.3	12.7	35.9	54.2	-	45M	50
ViDT-Swin-T [30]	MS training	36.3	56.3	37.8	16.4	39.0	54.3	-	29M	150
DAB-DETR-Swin-T [28]	MS training	45.2	66.8	47.8	24.2	49.0	64.8	100	47M	50
DN-DETR-Swin-T^∗	MS training	46.9	67.9	49.6	26.4	51.4	66.6	100	47M	50
Elastic-DETR-S	Ours ( $\tau_{max}=1.25$ )	45.9	66.8	49.3	22.9	49.9	66.5	78	48M	50
Elastic-DETR-M^†	Ours ( $\tau_{max}=1.50$ )	47.3	68.2	50.7	26.8	51.2	66.9	101	48M	50
Elastic-DETR-B^†	Ours ( $\tau_{max}=1.75$ )	48.3	69.1	51.9	28.3	52.5	66.7	143	48M	50
Elastic-DETR-L^†	Ours ( $\tau_{max}=2.00$ )	49.3	70.1	53.3	29.6	53.2	67.4	178	48M	50