VPD

This repo contains code of our paper "[Toward Accurate and Robust Pedestrian Detection via Variational Inference]"

Usage

Installation

1. Requirements

We have tested the following versions of OS and softwares:

• OS: Ubuntu 18.04.5 LTS
• CUDA: 10.1
• PyTorch 1.6.0
• Python 3.7.13

2. Install all other dependencies via:

pip install -r requirements.txt

Dataset

The CrowdHuman dataset can be downloaded from http://www.crowdhuman.org/. Orgnize the dataset folder in the following structure:

data
|-- CrowdHuman
    |-- Images
        |-- <name1>.<ImageFormat>
        |-- <name2>.<ImageFormat>
        ...
    |-- annotation_train.odgt
    |-- annotation_test.odgt

Train

1. Pretrain weights:

   Download the pretrained model here.

2. Config

   Edit config file in model/<method>/config.py, including dataset and network settings.

3. Run multi GPU distributed training:

   python tools/train.py -md <method>_fpn_<spd, spd+, vpd>

Evaluation

Run the evaluation by:

python tools/test.py -md <method>_fpn_<spd, spd+, vpd>.py -r <epoch_num>

The evaluation results can be found in model/<method>/outputs/eval_dump with epoch IDs (epoch_num, ..., 30).

Result

We offer the pre-trained weights on CrowdHuman datasets (Some unavailable model weights are coming soon):

Method	MR	AP	Weight
RetinaNet (SPD)	60.8%	84.9%
RetinaNet (SPD+)	60.5%	85.0%
RetinaNet (VPD)	56.2%	86.4%	retina_fpn_vpd
ATSS (SPD)	54.8%	85.4%
ATSS (SPD+)	54.0%	86.6%
ATSS (VPD)	52.1%	87.1%	atss_fpn_vpd
FreeAnchor (SPD)	51.8%	84.3%
FreeAnchor (SPD+)	51.1%	84.5%
FreeAnchor (VPD)	47.4%	84.7%	freeanchor_fpn_vpd

Acknowledgement

Our code is heavily based on Crowddet and mmdetection, thanks for their excellent work!