notesum.ai

Published at October 21

WildOcc: A Benchmark for Off-Road 3D Semantic Occupancy Prediction

cs.LG

cs.AI

Released Date: October 21, 2024

Authors: Heng Zhai, Jilin Mei, Chen Min, Liang Chen, Fangzhou Zhao, Yu Hu

Arxiv: https://arxiv.org/abs/2410.15792v1

Refer to caption

Method	Modality	IoU	mIoU	Grass	Tree	Bush	Puddle	Mud	Barrier	Rubble	Input Size	2D Backbone
SurroundOcc[3]	C	28.7	10.2	23.7	20.4	20.5	0.4	3.7	0.3	2.2	$900\times 1600$	R101-DCN
OccFormer[31]	C	27.2	10.5	24.3	21.1	20.7	0.5	2.3	0.7	3.6	$896\times 1600$	R101
C-CONet[2]	C	23.6	8.8	23.5	10.8	19.2	0.3	5.8	0.4	1.7	$896\times 1600$	R101
FB-Occ[1]	C	28.4	10.6	24.1	21.6	20.4	1.1	3.2	1.0	2.5	$896\times 1600$	R101
C-OFFOcc(Ours)	C	29.7	11.2	24.6	23.8	22.1	0.6	3.5	0.6	3.2	$896\times 1600$	R101
L-CONet[2]	L	32.4	11.0	25.3	26.7	23.4	0.1	1.4	0.9	0.4	-	-
M-CONet[2]	M	31.6	13.6	27.2	30.5	26.3	0.5	7.1	1.2	2.3	$896\times 1600$	R101
L-OFFOcc(Ours)	L	35.3	11.6	22.8	28.5	25.0	0.5	3.4	1.1	0.1	-	-
M-OFFOcc(Ours)	M	32.8	14.8	28.6	33.4	27.5	0.9	6.8	1.7	4.6	$896\times 1600$	R101
C-* denotes camera method, L-* denotes LiDAR method and M-* denotes camera-LiDAR fusion method.
C represents the modality is camera, L represents the modality is LiDAR and M represents the modality is camera-LiDAR fusion.
Underline means the best performance of camera method, and bold means the best performance of camera-LiDAR fusion method.