megvii-basedetection/lla

Official implementation of our paper "LLA: Loss-aware Label Assignment for Dense Pedestrian Detection" in Pytorch.

LLA: Loss-aware Label Assignment for Dense Pedestrian Detection

This project provides an implementation for "LLA: Loss-aware Label Assignment for Dense Pedestrian Detection" on PyTorch.

LLA is the first one-stage detector that surpasses two-stage detectors (e.g., Faster R-CNN) on CrowdHuman dataset. Experiments in the paper were conducted on the internal framework, thus we reimplement them on cvpods and report details as below.

Requirements

• cvpods

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/datasets
ln -s /path/to/your/crowdhuman/dataset crowdhuman

• Train & Test

git clone https://github.com/Megvii-BaseDetection/LLA.git
cd LLA/playground/detection/crowdhuman/lla.res50.fpn.crowdhuman.800size.30k  # 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 CrowdHuman val set

Model	Backbone	LR Sched.	Aux. Branch	NMS Thr.	MR	AP50	Recall	Download
FCOS	Res50	30k	CenterNess	0.6	54.4	86.0	94.1	weights
ATSS	Res50	30k	CenterNess	0.6	49.4	87.3	94.1	weights
Faster R-CNN	Res50	30k	-	0.5	48.5	84.3	87.1	weights
LLA.FCOS	Res50	30k	IoU	0.6	47.5	88.2	94.4	weights

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{ge2021lla,
  title={LLA: Loss-aware Label Assignment for Dense Pedestrian Detection},
  author={Ge, Zheng and Wang, Jianfeng and Huang, Xin and Liu, Songtao and Yoshie, Osamu},
  journal={arXiv preprint arXiv:2101.04307},
  year={2021}
}

lla's Issues

RetinaNet version of LLA?

Hi, thanks for sharing this excellent work, interesting and elegant idea! I have a question regarding the RetinaNet, do you plan to publish the LLA-RetinaNet code? Interested to do a comparison.

Results on cityperson dataset

Hi, thanks for sharing this excellent work! I have a question about the results on cityperson dataset. I trained the ATSS model on cityperson dataset and test on it with the code provided in cvpods/data/datasets/citypersons.py. I followed the same training setting as in Crowdhuman, as described in your paper. However I got unsatisfactory results (0.27 MR-2 on reasonable subset and 0.571 on HO subset) which has a huge gap with the results report in your paper. Do I miss something important? the config setting is as following: