zhangp14 / MVDet

Codes and MultiviewX dataset for ECCV 2020 paper: Multiview Detection with Feature Perspective Transformation

Multiview Detection with Feature Perspective Transformation [Website] [arXiv]

@inproceedings{hou2020multiview,
  title={Multiview Detection with Feature Perspective Transformation},
  author={Hou, Yunzhong and Zheng, Liang and Gould, Stephen},
  booktitle={ECCV},
  year={2020}
}

Overview

We release the PyTorch code for MVDet, a state-of-the-art multiview pedestrian detector; and MultiviewX dataset, a novel synthetic multiview pedestrian detection datatset.

Wildtrack	MultiviewX

Content

• MultiviewX dataset
  • Download MultiviewX
  • Build your own version
• MVDet Code
  • Dependencies
  • Data Preparation
  • Training

MultiviewX dataset

Using pedestrian models from PersonX, in Unity, we build a novel synthetic dataset MultiviewX.

MultiviewX dataset covers a square of 16 meters by 25 meters. We quantize the ground plane into a 640x1000 grid. There are 6 cameras with overlapping field-of-view in MultiviewX dataset, each of which outputs a 1080x1920 resolution image. We also generate annotations for 400 frames in MultiviewX at 2 fps (same as Wildtrack). On average, 4.41 cameras are covering the same location.

Download MultiviewX

Please refer to this link for download.

Build your own version

Please refer to this repo for a detailed guide & toolkits you might need.

MVDet Code

This repo is dedicated to the code for MVDet.

Dependencies

This code uses the following libraries

• python 3.7+
• pytorch 1.4+ & tochvision
• numpy
• matplotlib
• pillow
• opencv-python
• kornia
• matlab & matlabengine (required for evaluation) (see this link for detailed guide)

Data Preparation

By default, all datasets are in ~/Data/. We use MultiviewX and Wildtrack in this project.

Your ~/Data/ folder should look like this

Data
├── MultiviewX/
│   └── ...
└── Wildtrack/ 
    └── ...

Training

In order to train classifiers, please run the following,

CUDA_VISIBLE_DEVICES=0,1 python main.py -d wildtrack

This should automatically return evaluation results similar to the reported 88.2% MODA on Wildtrack dataset.