| Architecture | GroIE, ResNet, FPN, Global Context Block |
|---|---|
| lr sched | 1x |
| SHOW MORE |
| Architecture | Softmax, RPN, GroIE, Convolution, Dense Connections, FPN, ResNet, RoIAlign |
|---|---|
| lr sched | 1x |
| SHOW MORE |
| Architecture | ResNet, FPN, Global Context Block |
|---|---|
| lr sched | 1x |
| SHOW MORE |
| Architecture | ResNet, FPN, Global Context Block |
|---|---|
| lr sched | 1x |
| SHOW MORE |
| Architecture | FCN, RPN, Convolution, Dilated Convolution, FPN, Sigmoid Activation, ResNet, RoIAlign |
|---|---|
| lr sched | 1x |
| SHOW MORE |
| Architecture | Softmax, RPN, Convolution, Dense Connections, FPN, ResNet, RoIAlign |
|---|---|
| lr sched | 1x |
| SHOW MORE |
By Leonardo Rossi, Akbar Karimi and Andrea Prati from IMPLab.
We provide configs to reproduce the results in the paper for "A novel Region of Interest Extraction Layer for Instance Segmentation" on COCO object detection.
[ALGORITHM]
This paper is motivated by the need to overcome to the limitations of existing RoI extractors which select only one (the best) layer from FPN.
Our intuition is that all the layers of FPN retain useful information.
Therefore, the proposed layer (called Generic RoI Extractor - GRoIE) introduces non-local building blocks and attention mechanisms to boost the performance.
The results on COCO 2017 minival (5k images) are shown in the below table. You can find here the trained models.
| Backbone | Method | Lr schd | box AP | mask AP | Config | Download |
|---|---|---|---|---|---|---|
| R-50-FPN | Faster Original | 1x | 37.4 | config | model | log | |
| R-50-FPN | + GRoIE | 1x | 38.3 | config | model | log | |
| R-50-FPN | Grid R-CNN | 1x | 39.1 | config | model | log | |
| R-50-FPN | + GRoIE | 1x | config | |||
| R-50-FPN | Mask R-CNN | 1x | 38.2 | 34.7 | config | model | log |
| R-50-FPN | + GRoIE | 1x | 39.0 | 36.0 | config | model | log |
| R-50-FPN | GC-Net | 1x | 40.7 | 36.5 | config | model | log |
| R-50-FPN | + GRoIE | 1x | 41.0 | 37.8 | config | model | log |
| R-101-FPN | GC-Net | 1x | 42.2 | 37.8 | config | model | log |
| R-101-FPN | + GRoIE | 1x | config | model | log |
If you use this work or benchmark in your research, please cite this project.
@misc{rossi2020novel,
title={A novel Region of Interest Extraction Layer for Instance Segmentation},
author={Leonardo Rossi and Akbar Karimi and Andrea Prati},
year={2020},
eprint={2004.13665},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
The implementation of GROI is currently maintained by Leonardo Rossi.
Object Detection on COCO minival
| MODEL | BOX AP |
|---|---|
| GC-Net (R-101-FPN, 1x) | 42.2 |
| GC-Net GRoIE (R-50-FPN, 1x) | 41.0 |
| GC-Net (R-50-FPN, 1x) | 40.7 |
| Grid R-CNN (R-50-FPN, 1x) | 39.1 |
| Mask R-CNN GRoIE (R-50-FPN, 1x) | 39.0 |
| Faster R-CNN GRoIE (R-50-FPN, 1x) | 38.3 |
| Mask R-CNN (R-50-FPN, 1x) | 38.2 |
| Faster R-CNN (R-50-FPN, 1x) | 37.4 |
Instance Segmentation on COCO minival
| MODEL | MASK AP |
|---|---|
| GC-Net (R-101-FPN, 1x) | 37.8 |
| GC-Net GRoIE (R-50-FPN, 1x) | 37.8 |
| GC-Net (R-50-FPN, 1x) | 36.5 |
| Mask R-CNN GRoIE (R-50-FPN, 1x) | 36.0 |
| Mask R-CNN (R-50-FPN, 1x) | 34.7 |
