CenterNet2 Xingyi Zhou, Vladlen Koltun, Philipp Krhenbhl UT Austin - - PowerPoint PPT Presentation

▶
centernet2
SMART_READER_LITE
LIVE PREVIEW

CenterNet2 Xingyi Zhou, Vladlen Koltun, Philipp Krhenbhl UT Austin - - PowerPoint PPT Presentation

Feb 28, 2024 754 likes •908 views

CenterNet2 Xingyi Zhou, Vladlen Koltun, Philipp Krhenbhl UT Austin & Intel Labs 1 Conventional two-stage detector Backbone Classifier BB regression ROIAlign Stage 1 Stage 2 45ms 8ms Ren et. al, Faster R-CNN: Towards

slide-1
SLIDE 1

CenterNet2

Xingyi Zhou, Vladlen Koltun, Philipp Krähenbühl UT Austin & Intel Labs

1

slide-2
SLIDE 2

Conventional two-stage detector

Backbone

…

Classifier BB regression

ROIAlign

Stage 1 45ms Stage 2 8ms

Ren et. al, Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, NIPS 2015

slide-3
SLIDE 3

Cascade detector

Backbone

…

Classifier BB regression

ROIAlign

Stage 1 45ms Stage 2 8ms Stage 3 8ms

ROIAlign …

Classifier BB regression

… … Cai et. al, Cascade R-CNN: Delving into High Quality Object Detection, CVPR 2018

slide-4
SLIDE 4

One-stage detector

Backbone Classifier BB regression Stage 1 53ms

Lin et. al, Focal Loss for Dense Object Detection, ICCV 2017

slide-5
SLIDE 5

CenterNet

Backbone keypoint detection size regression Stage 1 53ms

Zhou et. al, Objects as Points, arXiv 2019

slide-6
SLIDE 6

CenterNet2

Backbone keypoint detection size regression Stage 1 51ms

…

Classifier BB regression

ROIAlign

Stage 2 2ms Stage 3 2ms

… ROIAlign

Classifier BB regression

slide-7
SLIDE 7

Results

7

COCO box mAP

38 39.75 41.5 43.25 45 CenterNet-FPN CascadeRCNN CenterNet2

42.9 42.1 39.6 COCO runtime (ms)

20 40 60 80 CenterNet-FPN CascadeRCNN CenterNet2

60 70 53 LVIS box mAP

20 22.5 25 27.5 30 CascadeRCNN CenterNet2

26.9 24

  • Res50, Multi-scale, 1x
slide-8
SLIDE 8

Federated datasets

8

… Negatives Positives … Unlabeled …

Gupta et. al, LVIS: A Dataset for Large Vocabulary Instance Segmentation, CVPR 2019

slide-9
SLIDE 9

Cross entropy

9

… Positives … … Negatives

slide-10
SLIDE 10

Equalization loss (Tan et al. 2020)

  • For frequent classes
  • Sigmoid cross entropy
  • For rare classes
  • Ignore negative loss from

foreground

10

Tan et. al, Equalization Loss for Long-Tailed Object Recognition CVPR 2020

slide-11
SLIDE 11

Federated Loss (ours)

  • Positive classes
  • Sigmoid cross entropy
  • Negative or unlabeled classes
  • Sampled to apply negative

loss

  • Sample based on frequency

11

slide-12
SLIDE 12

Results

7 14 21 28 35 Box AP Box APr Box APc Box APf

33.3 26 16.1 27.1 31.5 24.6 15.5 25.7 31.5 21.9 8.2 23.3 32.7 22.9 7.6 24

Softmax-CE Sigmoid-CE EQL FedLoss

12

slide-13
SLIDE 13

Challenge model

mAP 24 28 32 36 40 CenterNet2 +Mask +2x +FPN2-6 +X101 +DCN +PointRend +Larger input +Test-aug

27.2 36.1 34.9 34 32.1 30.3 28.2 27.4 25.3 38.5 37.3 36.7 35.9 33.9 31.5 30.6 28.6 28.2

Box mAP Mask mAP Official baseline

13

slide-14
SLIDE 14

Thanks!