Computer Vision News - January 2019

The validation bbox AP curves are shown side by side below: using GN with ResNet-101 (R101) backbone and using SyncBN with ResNet-50 (R50) backbone. Each figure compares the curves for models trained from random initialization vs. ImageNet pre-trained then fine-tuned. The figure in the next page presents comparisons between training from random initialization vs. pre-training then fine-tuning on various systems using Mask R- CNN, including: 1) baselines using FPN and GN, 2) baselines with training time multi-scale augmentation, 3) baselines with Cascade RCNN and training-time augmentation and 4) plus test-time multi-scale augmentation - left: R50; right: R101. Rethinking ImageNet Pre-training 7 Research Computer Vision News