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 AI Center Moscow
https://arxiv.org/abs/2005.05708

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:

• AdaptIS ToyV1 and ToyV2
• CrowdHuman
• WiderPerson

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

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:

@article{rukhovich2020iterdet,
  title={IterDet: Iterative Scheme for Object Detection in Crowded Environments},
  author={Danila Rukhovich, Konstantin Sofiiuk, Danil Galeev, Olga Barinova, Anton Konushin},
  journal={arXiv preprint arXiv:2005.05708},
  year={2020}
}