FoveaBox: Beyond Anchor-based Object Detector

8 Apr 2019 Tao Kong Fuchun Sun Huaping Liu Yuning Jiang Lei LI Jianbo Shi

We present FoveaBox, an accurate, flexible, and completely anchor-free framework for object detection. While almost all state-of-the-art object detectors utilize predefined anchors to enumerate possible locations, scales and aspect ratios for the search of the objects, their performance and generalization ability are also limited to the design of anchors... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Object Detection COCO minival FoveaBox (ResNet-50-FPN, 600x600) box AP 36.0 # 60
AP50 55.2 # 46
AP75 37.9 # 48
APS 18.6 # 40
APM 39.4 # 43
APL 50.5 # 40
Object Detection COCO minival FoveaBox (ResNet-101-FPN, 800x800) box AP 38.9 # 50
AP50 58.4 # 40
AP75 41.5 # 39
APS 22.3 # 33
APM 43.5 # 31
APL 51.7 # 36
Object Detection COCO minival FoveaBox+Retina (ResNet-50) box AP 38.1 # 54
AP50 57.8 # 43
AP75 40.5 # 44
Object Detection COCO minival FoveaBox (ResNet-101-FPN, 600x600) box AP 38 # 55
AP50 57.8 # 43
AP75 40.2 # 45
APS 19.5 # 39
APM 42.2 # 34
APL 52.7 # 31
Object Detection COCO test-dev FoveaBox (ResNeXt-101) box AP 43.9 # 48
AP50 63.5 # 48
AP75 47.7 # 52
APS 26.8 # 43
APM 46.9 # 49
APL 55.6 # 47
box AP 42.1 # 58
AP50 61.9 # 58
AP75 45.2 # 64
APM 46.8 # 50
APS 24.9 # 54

Methods used in the Paper