AI Paper Atlas

Research map / Detection and segmentation / CNN features

Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

Chen Sun, Abhinav Shrivastava, Saurabh Singh, Abhinav Gupta

ICCV 2017 · Detection and segmentation › CNN features · approach: Convolution-transformer backbones · 2,652 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
pretrain on JFT-300M then fine-tune instead of random initResNet-101 · ImageNet valtop-1 accuracy: 77.5 -> 79.2Improved
pretrain on JFT-300M instead of ImageNetResNet-101 Faster RCNN · COCO test-dev detectionmAP@[0.5,0.95]: 34.3 -> 36.7Improved
add ImageNet initialization to JFT-300M pretrainingResNet-101 Faster RCNN · COCO test-dev detectionmAP@[0.5,0.95]: 36.7 -> 37.4Improved
pretrain on JFT-300M instead of ImageNetResNet-101 Faster RCNN · PASCAL VOC 2007 test detectionmAP@0.5: 76.3 -> 81.4Improved
add ImageNet initialization to JFT-300M pretrainingResNet-101 Faster RCNN · PASCAL VOC 2007 test detectionmAP@0.5: 81.4 -> 81.3No clear effect
pretrain on JFT-300M instead of ImageNetDeepLab-ASPP-L · PASCAL VOC 2012 val segmentationmIoU: 73.6 -> 75.3Improved
add ImageNet initialization to JFT-300M pretrainingDeepLab-ASPP-L · PASCAL VOC 2012 val segmentationmIoU: 75.3 -> 76.5Improved
pretrain on JFT-300M instead of ImageNetG-RMI pose ResNet-101 · COCO test-dev pose estimationAP: 62.4 -> 64.8Improved
add ImageNet initialization to JFT-300M pretrainingG-RMI pose ResNet-101 · COCO test-dev pose estimationAP: 64.8 -> 64.4No clear effect
pretrain JFT-300M for more epochs (1.3 to 4) (ablation)ResNet-101 Faster RCNN · COCO minivalmAP@[0.5,0.95]: 35.0 -> 36.8Improved
pretrain ImageNet for more epochs (3 to 150) (ablation)ResNet-101 Faster RCNN · COCO minivalmAP@[0.5,0.95]: 22.2 -> 34.5Improved
use 18K JFT classes instead of 1K ImageNet-matching classes (ablation)30M-image JFT pretraining · COCO minivalmAP@[0.5,0.95]: 31.2 -> 31.9No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy