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

CenterNet2 Xingyi Zhou, Vladlen Koltun, Philipp Krähenbühl 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 Real-Time Object Detection with Region Proposal Networks, NIPS 2015

Cascade detector Backbone Classifier BB regression ROIAlign … Classifier … BB regression ROIAlign … Stage 1 Stage 2 Stage 3 … 45ms 8ms 8ms Cai et. al, Cascade R-CNN: Delving into High Quality Object Detection, CVPR 2018

One-stage detector Classifier Backbone BB regression Stage 1 53ms Lin et. al, Focal Loss for Dense Object Detection, ICCV 2017

CenterNet keypoint detection Backbone size regression Stage 1 53ms Zhou et. al, Objects as Points, arXiv 2019

CenterNet2 keypoint detection Backbone size regression Classifier BB regression ROIAlign Classifier … BB regression ROIAlign Stage 1 Stage 2 Stage 3 … 51ms 2ms 2ms

Results • Res50, Multi-scale, 1x COCO runtime (ms) COCO box mAP LVIS box mAP 80 70 60 45 30 53 60 42.9 26.9 43.25 27.5 42.1 40 24 41.5 25 39.6 20 39.75 22.5 38 0 20 CenterNet-FPN CascadeRCNN CenterNet2 CenterNet-FPN CascadeRCNN CenterNet2 CascadeRCNN CenterNet2 7

Federated datasets Positives Negatives Unlabeled … … … 8 Gupta et. al, LVIS: A Dataset for Large Vocabulary Instance Segmentation, CVPR 2019

Cross entropy Positives Negatives … … … 9

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

Federated Loss (ours) • Positive classes • Sigmoid cross entropy • Negative or unlabeled classes • Sampled to apply negative loss • Sample based on frequency 11

Results Softmax-CE Sigmoid-CE EQL FedLoss 33.3 35 32.7 31.5 31.5 27.1 26 28 25.7 24.6 24 23.3 22.9 21.9 21 16.1 15.5 14 8.2 7.6 7 Box AP Box APr Box APc Box APf 12

Challenge model Box mAP Mask mAP O ffi cial baseline 40 38.5 37.3 36.7 36.1 35.9 36 34.9 34 33.9 32.1 mAP 31.5 32 30.6 30.3 28.6 28.2 28.2 27.4 28 27.2 25.3 24 CenterNet2 +Mask +2x +FPN2-6 +X101 +DCN +PointRend +Larger input +Test-aug 13

Thanks!