FCOS: Fully Convolutional One-Stage Object Detection

We propose a fully convolutional one-stage object detector (FCOS) to solve object detection in a per-pixel prediction fashion, analogue to semantic segmentation. Almost all state-of-the-art object detectors such as RetinaNet, SSD, YOLOv3, and Faster R-CNN rely on pre-defined anchor boxes... (read more)

PDF Abstract ICCV 2019 PDF ICCV 2019 Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Object Detection COCO minival FCOS (ResNet-50-FPN + improvements) box AP 38.6 # 51
AP50 57.4 # 44
AP75 41.4 # 40
APS 22.3 # 33
APM 42.5 # 33
APL 49.8 # 41
Object Detection COCO test-dev FCOS (ResNeXt-32x8d-101-FPN) box AP 42.7 # 55
AP50 62.2 # 56
AP75 46.1 # 60
APS 26.0 # 48
APM 45.6 # 56
APL 52.6 # 61
Object Detection COCO test-dev FCOS (HRNet-W32-5l) box AP 42.0 # 59
AP50 60.4 # 63
AP75 45.3 # 63
APS 25.4 # 51
APM 45.0 # 59
APL 51.0 # 70
Object Detection COCO test-dev FCOS (ResNeXt-101-64x4d-FPN) box AP 43.2 # 52
AP50 62.8 # 53
AP75 46.6 # 56
APS 26.5 # 45
APM 46.2 # 55
APL 53.3 # 58
Object Detection COCO test-dev FCOS (ResNeXt-64x4d-101-FPN 4 + improvements) box AP 44.7 # 44
AP50 64.1 # 44
AP75 48.4 # 49
APS 27.6 # 38
APM 47.5 # 45
APL 55.6 # 47

Methods used in the Paper