/iterdet

[S+SSPR2020] IterDet: Iterative Scheme for Object Detection in Crowded Environments

Primary LanguagePythonMozilla Public License 2.0MPL-2.0

PWC PWC

IterDet: Iterative Scheme for Object Detection in Crowded Environments

This project hosts the code for implementing the IterDet scheme for object detection, as presented in our paper:

IterDet: Iterative Scheme for Object Detection in Crowded Environments
Danila Rukhovich, Konstantin Sofiiuk, Danil Galeev, Olga Barinova, Anton Konushin
Samsung Research
https://arxiv.org/abs/2005.05708

drawing

Installation

This implementation is based on mmdetection framework.

All our modifications against their `v2.0.0` release are listed below:
  • configs/iterative/*
  • demo/iterative/*
  • mmdet/datasets/__init__.py
  • mmdet/datasets/pipelines/transforms.py
  • mmdet/datasets/pipelines/formating.py
  • mmdet/datasets/crowd_human.py
  • mmdet/models/dense_heads/anchor_head.py
  • mmdet/models/dense_heads/rpn_head.py
  • mmdet/models/roi_heads/bbox_heads/bbox_head.py
  • mmdet/models/backbones/resnet.py
  • mmdet/models/detectors/__init__.py
  • mmdet/models/detectors/iterdet_faster_rcnn.py
  • mmdet/models/detectors/iderdet_retinanet.py
  • tools/convert_datasets/crowd_human.py
  • tools/convert_datasets/toy.py
  • tools/convert_datasets/wider_person.py
  • requirements/runtime.txt
  • docker/Dockerfile

Please refer to original install.md for installation. Do not forget to update the original github repository link, and install requirements.txt. For v1.2.0 release follow v1 branch.

Config files and tools for converting annotations to COCO format are provided for the following datasets:

Get Started

Please see original getting_started.md for the basic usage examples. Iterdet configs can be used for train and test scripts:

bash tools/dist_train.sh configs/iterdet/crowd_human_full_faster_rcnn_r50_fpn_2x.py 8 --validate
bash tools/dist_test.sh configs/iterdet/crowd_human_full_faster_rcnn_r50_fpn_2x.py \
    work_dirs/iterdet/crowd_human_full_faster_rcnn_r50_fpn_2x/latest.pth 8

Models

State-of-the-art models for all datasets are trained on top of Faster RCNN based on ResNet-50. Metrics are given for 2 iterations IterDet inference.

Dataset Download Link Recall AP mMR
AdaptIS Toy V1 toy_v1.pth 99.60 99.25
AdaptIS Toy V2 toy_v2.pth 99.29 99.00
CrowdHuman (full) crowd_human_full.pth 95.80 88.08 49.44
CrowdHuman (visible) crowd_human_visible.pth 91.63 85.33 55.61
WiderPerson wider_person.pth 97.15 91.95 40.78

Example Detections

drawing

Examples of IterDet results on ToyV1, ToyV2, CrowdHuman (with full body annotataions), and WiderPerson. The boxes found on the first and second iterations are marked in green and yellow respectively.

License

The code is released under the MPL 2.0 License. MPL is a copyleft license that is easy to comply with. You must make the source code for any of your changes available under MPL, but you can combine the MPL software with proprietary code, as long as you keep the MPL code in separate files.

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{rukhovich2021iterdet,
  title={IterDet: Iterative Scheme for Object Detection in Crowded Environments},
  author={Danila Rukhovich, Konstantin Sofiiuk, Danil Galeev, Olga Barinova, Anton Konushin},
  booktitle={Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR International Workshops, S+ SSPR 2020, Padua, Italy, January 21--22, 2021, Proceedings},
  pages={344},
  organization={Springer Nature}
}