hli2020 / Feature_intertwiner

Feature Intertwiner for Object Detection

A PyTorch implementation of our paper published at ICLR 2019.

By Hongyang Li, Bo Dai, Shaoshuai Shi, Wanli Ouyang, and Xiaogang Wang.

Paper: [arXiv] [Openreview]

A 50-min talk presented at GTC 2019: [GTC Video] [GTC Slides]

Overview

Our assumption is that semantic features for one category should be the same as shown in (a) below. Due to the inferior up-sampling design in RoI operation, shown in (b), the reliable set (green) could guide the feature learning of the less reliable set (blue).

Here comes the proposed feature intertwiner:

• PyTorch 0.3
• Code/framework based on Mask-RCNN.
• Datasets: COCO and Pascal VOC (not in this repo)

How to run

Follow instructions in INSTALL.md to set up datasets, symlinks, compilation, etc.

To train, sh script/base_4gpu 105/meta_105_quick_1 0,2,5,7 # gpu ids or:

simply execute python main.py. The configurations are stored in the configs folder.

To test, change the flag --phase in main.py to inference.

Performance

Object detection single-model performance (bounding box AP) on the COCO test-dev. The InterNet multi-scale is achieved with data augmentation, 1.5× longer training time and multi-scale training. Our InterNet is also a two-stage detector.

Adapting Feature Intertwiner to your own task

This is probably the most concerned part for most audience.

Citation

Please cite in the following manner if you find it useful in your research:

@inproceedings{li2019_internet,
  title = {{Feature Intertwiner for Object Detection}},
  author = {Hongyang Li and Bo Dai and Shaoshuai Shi and Wanli Ouyanbg and Xiaogang Wang},
  booktitle = {ICLR},
  year = {2019}
}