Title: Light-Head R-CNN: In Defense of Two-Stage Object Detector
Contribution:
- proposed a 2-stage detector with good accuracy and promising speed compare with single-stage detector
- investigate the problems with Faster-RCNN (global avg pooling harmful for spatial loc with out sharing compution) & RFCN (with a large score map for ROI pooling which is costly)
- proposed thin feature maps for generating small channel ROI feature maps, improving accuracy, save mem/ computation
- detailed hyper-param setting & experiments give strong results, also show techs which improved mAP
Experiment:
- evaluated on COCO
- adopt dilated conv & OHEM
- R-FCN as baseline
- with ResNet as backbone, achieve 41.5 mmAP
- with Xception achieve 30.7 mmAP, 102 FPS