Cascade RCNN Delving into High Quality Object Detection

Cascade R-CNN: Delving into High Quality Object Detection Zhaowei Cai, Nuno Vasconcelos University of California San Diego SVCL 1

Object Detection • Definition of positive/negative samples – Ambiguous!! positive? ? SVCL 2

Object Detection • Standard definition – is to use an intersection over union (Io. U) threshold (usually 0. 5). positive √ positive × SVCL 3

Object Detection • During training – these samples are on opposite sides of the class boundary. Io. U=0. 5 SVCL 4

Object Detection • However, this threshold is loose – many positive samples are of low quality. Io. U=0. 5 low quality positives SVCL 5

Current Object Detection • Detection results can be quite noisy! low quality detector SVCL 6

High Quality Object Detection • Q: how to train a high quality object detector? high quality detector SVCL 7

Low Quality Object Detection • Training with low Io. U threshold Io. U=0. 5 SVCL 8

High Quality Object Detection • Training with high Io. U threshold Io. U=0. 7 SVCL 9

High Quality Object Detection Two problems 1. Training overfitting, due to exponentially vanishing positive samples. 2. Inference-time quality mismatch between the detector and its input proposals. SVCL 10

Cascade R-CNN • Multi-stage framework, increasing Io. U threshold Io. U=0. 5 Faster R-CNN Io. U=0. 6 Io. U=0. 7 Cascade R-CNN SVCL 11

Why Cascade R-CNN? Two benefits 1. Reduce training overfitting, since the proposal quality is sequentially improved. 2. Reduce Inferencetime quality mismatch, because the same cascade is also used at inference. Io. U=0. 5 SVCL Io. U=0. 6 Io. U=0. 7 12

Effective High Quality Detection • Large gains for high quality detection 60 40 baseline cascade 20 0 AP 50 AP 60 AP 70 SVCL AP 80 AP 90 Detection Quality 13

State-of-the-art Results • The best single model results on COCO RCN N de Ne t Ca sc a At tr ac tio -C NN N as k. R FC rm ab fo De M le G- R- RM I N FP + NN ++ t -C r. R Fa st e Re t in a. N e D SS YO L O 45 40 35 30 25 20 15 10 5 0 SVCL 14

Welcome to our poster! (P 2 -3 -D 14) Cascade R-CNN: Delving into High Quality Object Detection SVCL 15