Deep Structured Instance Graph for Distilling Object Detectors

Authors: Yixin Chen, Pengguang Chen, Shu Liu, Liwei Wang, Jiaya Jia.

[pdf] [arXiv] [supp] [bibtex]

This repo provides the implementation of paper "Deep Structured Instance Graph for Distilling Object Detectors"(Dsig) based on detectron2. Specifically, aiming at solving the feature imbalance problem while further excavating the missing relation inside semantic instances, we design a graph whose nodes correspond to instance proposal-level features and edges represent the relation between nodes. We achieve new state-of-the-art results on the COCO object detection task with diverse student-teacher pairs on both one- and two-stage detectors.

• Python >= 3.6
• Pytorch >= 1.7.0
• Torchvision >= 0.8.1
• Pycocotools 2.0.2

Follow the install instructions in detectron2, note that in this repo we use detectron2 commit version ff638c931d5999f29c22c1d46a3023e67a5ae6a1. Download COCO dataset and export DETECTRON2_DATASETS=$COCOPATH to direct to COCO dataset. We prepare our pre-trained weights for training in Student-Teacher format, please follow the instructions in Pretrained.

We prepare training configs following the detectron2 format. For training a Faster R-CNN R18-FPN student with a Faster R-CNN R50-FPN teacher on 4 GPUs:

./start_train.sh train projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

For testing:

./start_train.sh eval projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

For debugging:

./start_train.sh debugtrain projects/Distillation/configs/Distillation-FasterRCNN-R18-R50-dsig-1x.yaml

Faster R-CNN:

RetinaNet:

More models and results will be released soon.

@InProceedings{Chen_2021_ICCV,
    author    = {Chen, Yixin and Chen, Pengguang and Liu, Shu and Wang, Liwei and Jia, Jiaya},
    title     = {Deep Structured Instance Graph for Distilling Object Detectors},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {4359-4368}
}

Please contact yxchen@cse.cuhk.edu.hk.