End-to-End Object Detection with Fully Convolutional Network

This project provides an implementation for "End-to-End Object Detection with Fully Convolutional Network" on PyTorch.

Experiments in the paper were conducted on the internal framework, thus we reimplement them on cvpods and report details as below.

Requirements

• cvpods
• scipy >= 1.5.4

Get Started

• install cvpods locally (requires cuda to compile)

python3 -m pip install 'git+https://github.com/Megvii-BaseDetection/cvpods.git'
# (add --user if you don't have permission)

# Or, to install it from a local clone:
git clone https://github.com/Megvii-BaseDetection/cvpods.git
python3 -m pip install -e cvpods

# Or,
pip install -r requirements.txt
python3 setup.py build develop

• prepare datasets

cd /path/to/cvpods
cd datasets
ln -s /path/to/your/coco/dataset coco

• Train & Test

git clone https://github.com/Megvii-BaseDetection/DeFCN.git
cd DeFCN/playground/detection/coco/poto.res50.fpn.coco.800size.3x_ms  # for example

# Train
pods_train --num-gpus 8

# Test
pods_test --num-gpus 8 \
    MODEL.WEIGHTS /path/to/your/save_dir/ckpt.pth # optional
    OUTPUT_DIR /path/to/your/save_dir # optional

# Multi node training
## sudo apt install net-tools ifconfig
pods_train --num-gpus 8 --num-machines N --machine-rank 0/1/.../N-1 --dist-url "tcp://MASTER_IP:port"

Results on COCO2017 val set

model	assignment	with NMS	lr sched.	mAP	mAR	download
FCOS	one-to-many	Yes	3x + ms	41.4	59.1	weight | log
FCOS baseline	one-to-many	Yes	3x + ms	40.9	58.4	weight | log
Anchor	one-to-one	No	3x + ms	37.1	60.5	weight | log
Center	one-to-one	No	3x + ms	35.2	61.0	weight | log
Foreground Loss	one-to-one	No	3x + ms	38.7	62.2	weight | log
POTO	one-to-one	No	3x + ms	39.2	61.7	weight | log
POTO + 3DMF	one-to-one	No	3x + ms	40.6	61.6	weight | log
POTO + 3DMF + Aux	mixture*	No	3x + ms	41.4	61.5	weight | log

* We adopt a one-to-one assignment in POTO and a one-to-many assignment in the auxiliary loss, respectively.

• 2x + ms schedule is adopted in the paper, but we adopt 3x + ms schedule here to achieve higher performance.
• It's normal to observe ~0.3AP noise in POTO.

Results on CrowdHuman val set

model	assignment	with NMS	lr sched.	AP50	mMR	recall	download
FCOS	one-to-many	Yes	30k iters	86.1	54.9	94.2	weight | log
ATSS	one-to-many	Yes	30k iters	87.2	49.7	94.0	weight | log
POTO	one-to-one	No	30k iters	88.5	52.2	96.3	weight | log
POTO + 3DMF	one-to-one	No	30k iters	88.8	51.0	96.6	weight | log
POTO + 3DMF + Aux	mixture*	No	30k iters	89.1	48.9	96.5	weight | log

* We adopt a one-to-one assignment in POTO and a one-to-many assignment in the auxiliary loss, respectively.

• It's normal to observe ~0.3AP noise in POTO, and ~1.0mMR noise in all methods.

Ablations on COCO2017 val set

model	assignment	with NMS	lr sched.	mAP	mAR	note
POTO	one-to-one	No	6x + ms	40.0	61.9
POTO	one-to-one	No	9x + ms	40.2	62.3
POTO	one-to-one	No	3x + ms	39.2	61.1	replace Hungarian algorithm by argmax
POTO + 3DMF	one-to-one	No	3x + ms	40.9	62.0	remove GN in 3DMF
POTO + 3DMF + Aux	mixture*	No	3x + ms	41.5	61.5	remove GN in 3DMF

* We adopt a one-to-one assignment in POTO and a one-to-many assignment in the auxiliary loss, respectively.

• For one-to-one assignment, more training iters lead to higher performance.
• The argmax (also known as top-1) operation is indeed the approximate solution of bipartite matching in dense prediction methods.
• It seems harmless to remove GN in 3DMF, which also leads to higher inference speed.

Acknowledgement

This repo is developed based on cvpods. Please check cvpods for more details and features.

License

This repo is released under the Apache 2.0 license. Please see the LICENSE file for more information.

Citing

If you use this work in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@article{wang2020end,
  title   =  {End-to-End Object Detection with Fully Convolutional Network},
  author  =  {Wang, Jianfeng and Song, Lin and Li, Zeming and Sun, Hongbin and Sun, Jian and Zheng, Nanning},
  journal =  {arXiv preprint arXiv:2012.03544},
  year    =  {2020}
}

Contributing to the project

Any pull requests or issues about the implementation are welcome. If you have any issue about the library (e.g. installation, environments), please refer to cvpods.