Learning Object Bounding Boxes for 3D Instance Segmentation on Point - - PowerPoint PPT Presentation

▶
learning object bounding boxes for 3d instance
SMART_READER_LITE
LIVE PREVIEW

Learning Object Bounding Boxes for 3D Instance Segmentation on Point - - PowerPoint PPT Presentation

Mar 10, 2024 228 likes •346 views

Neural Information Processing Systems (NeurIPS) 2019 Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds B. Yang, J. Wang, R. Clark, Q. Hu, S. Wang, A. Markham, N. Trigoni Background: 2D Instance Segmentation 3D

slide-1
SLIDE 1

Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds

  • B. Yang,
  • J. Wang,
  • R. Clark,
  • Q. Hu,
  • S. Wang, A. Markham, N. Trigoni

Neural Information Processing Systems (NeurIPS) 2019

slide-2
SLIDE 2

Background:

3D Point Cloud

2D Instance Segmentation

Mask RCNN

3D Instance Segmentation

?

slide-3
SLIDE 3

Background:

Proposal-free Proposal-based

Ø Low objectness Ø Heavy post-processing (grouping)

SGPN (CVPR’18); ASIS (CVPR’19); JSIS3D (CVPR’19); 3D-BEVIS (GCPR '19); MTML (ICCV’19); MASC (arXiv’19)

Limitations

Ø Two-stage training Ø Heavy post-processing (NMS)

3D-SIS (CVPR’19); GSPN (CVPR’19)

slide-4
SLIDE 4

Our Method (3D-BoNet):

Ø Each object is uniquely detected and segmented. Ø The learnt 3D bounding boxes guarantee high objectness. Ø It’s end-to-end trainable, no post-processing, and efficient.

Highlights of our pipeline

slide-5
SLIDE 5

Our Method (3D-BoNet):

Input Point Cloud

§ Predicted bounding boxes § Predicted bbox scores ü GT bounding boxes ü GT bbox scores

Optimal Association

0.95 0.83 0.98 0.35 0.07 1.0 1.0 1.0

slide-6
SLIDE 6

Our Method (3D-BoNet): Multiple criteria to match a pred bbox with a GT bbox

slide-7
SLIDE 7

Our Method (3D-BoNet): End-to-end training losses

slide-8
SLIDE 8

Quantitative Results:

slide-9
SLIDE 9

Qualitative Results:

slide-10
SLIDE 10

Intermediate Results:

slide-11
SLIDE 11

Thank You