Codes in this folder is an implementation of InstaBoost for detectron of this version.
Prepare and run detectron according to ORIREADME.md.
Since loading images and packaging annotations into roidb are in different stages in this code, implementing InstaBoost is a little complex.
Main codes are in lib/roi_data/minibatch.py. These codes mainly do three works, fetch image and annotations, InstaBoost, and package annotations into roidb.
First two steps are easy to understand, while the third step might be confusing. We try to reuse combined_roidb_for_training function provided in lib/datasets/roidb.py to get new roidbs which are consistent with the previous ones. Therefore, we need to passing parameters img_id, new_ann, coco to infer a certain image and annotaion of the coco-style dataset we use before. We make a copy of lib/dataset folder, called lib/datasetAug, and changes some codes. Details are shown below.
Line 8 - line 11 in lib/roi_data/minibatch.py to import packages and load coco annotations.
Line 37, line 42, line 43 and line 63 in lib/datasetsAug/roidb.py to pass parameter coco.
Line 56, line 70, line 89 - line 90, line 136 - line 142, line 147 - line 171 and line 222 - line 223 in lib/datasetsAug/json_dataset.py to pass parameter coco, img_id, new_ann to replace these variables in original code.
4x configurations are available in InstaBoost_configs.
For your conveinience of evaluation and comparison, we report the evaluation number on COCO val below. In our paper, the numbers are obtained from test-dev.
| InstaBoost | Network | Backbone | Lr schd | box AP | mask AP | Download |
|---|---|---|---|---|---|---|
| × | Mask R-CNN | R-50-FPN | 1x | 37.7 | 33.7 | - |
| × | Mask R-CNN | R-50-FPN | 2x | 38.7 | 34.6 | - |
| × | Mask R-CNN | R-50-FPN | 3x | 38.0 | 33.9 | Baidu / Google |
| √ | Mask R-CNN | R-50-FPN | 4x | 40.3 | 35.7 | Baidu / Google |
| × | Mask R-CNN | R-101-FPN | 1x | 40.0 | 35.9 | - |
| √ | Mask R-CNN | R-101-FPN | 4x | 42.8 | 37.5 | Baidu / Google |
If you use this toolbox or benchmark in your research, please cite this project.
@article{Fang2019InstaBoost,
author = {Fang, Hao-Shu and Sun, Jianhua and Wang, Runzhong and Gou, Minghao and Li, Yong-Lu and Lu, Cewu},
title = {InstaBoost: Boosting Instance Segmentation Via Probability Map Guided Copy-Pasting},
journal={arXiv preprint arXiv:1908.07801},
year = {2019}
}