Focal Loss for Dense Object Detection

ICCV 2017 Tsung-Yi LinPriya GoyalRoss GirshickKaiming HePiotr Dollár

The highest accuracy object detectors to date are based on a two-stage approach popularized by R-CNN, where a classifier is applied to a sparse set of candidate object locations. In contrast, one-stage detectors that are applied over a regular, dense sampling of possible object locations have the potential to be faster and simpler, but have trailed the accuracy of two-stage detectors thus far... (read more)

PDF Abstract

Code


Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT LEADERBOARD
Real-Time Object Detection COCO FPN R-101 @256 MAP 37.8 # 6
APM 41.1 # 1
Region Proposal COCO test-dev RPN+Focal Loss AR100 50.2 # 3
AR1000 60.9 # 3
ARL 67.5 # 3
ARM 58.2 # 3
ARS 33.9 # 2
AR300 56.6 # 2
Object Detection COCO test-dev RetinaNet (ResNet-101-FPN) box AP 39.1 # 59
AP50 59.1 # 56
AP75 42.3 # 56
APS 21.8 # 57
APM 42.7 # 50
APL 50.2 # 58
Object Detection COCO test-dev RetinaNet (ResNeXt-101-FPN) box AP 40.8 # 48
AP50 61.1 # 46
AP75 44.1 # 48
APS 24.1 # 43
APM 44.2 # 44
APL 51.2 # 53

Results from Other Papers


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK SOURCE PAPER COMPARE
Object Counting CARPK RetinaNet (2018) MAE 24.58 # 5
RMSE 33.12 # 3
Dense Object Detection SKU-110K RetinaNet AP .455 # 3
AP75 .389 # 2
Face Identification Trillion Pairs Dataset F-Softmax Accuracy 39.80 # 5
Face Verification Trillion Pairs Dataset F-Softmax Accuracy 37.14 # 5