Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

The success of deep learning in vision can be attributed to: (a) models with high capacity; (b) increased computational power; and (c) availability of large-scale labeled data. Since 2012, there have been significant advances in representation capabilities of the models and computational capabilities of GPUs... (read more)

PDF Abstract ICCV 2017 PDF ICCV 2017 Abstract
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Object Detection COCO test-dev Faster R-CNN (ImageNet+300M) box AP 37.4 # 77
AP50 58 # 71
AP75 40.1 # 76
APS 17.5 # 78
APM 41.1 # 69
APL 51.2 # 64
Pose Estimation COCO test-dev Faster R-CNN (ImageNet+300M) AP 64.4 # 9
AP50 85.7 # 8
AP75 70.7 # 9
APL 69.8 # 10
APM 61.8 # 6
Image Classification ImageNet ResNet-101 (JFT-300M Finetuning) Top 1 Accuracy 79.2% # 75
Top 5 Accuracy 94.7% # 48
Semantic Segmentation PASCAL VOC 2007 DeepLabv3 (ImageNet+300M) Mean IoU 81.3 # 2
Semantic Segmentation PASCAL VOC 2012 val DeepLabv3 (ImageNet+300M) mIoU 76.5% # 13

Methods used in the Paper