MultiviewC engine and dataset toolkits!

MultiviewC engine based on UE4.

MultiviewC engine, the platform that support for multi-person online data collection, are released!

Data annotation

The engine captures accurate 3D and 2D information of the target (cattle), as well as the target's movements and the spatial distribution of the farm. Among other things, the cattle have a random running logic where they naturally stop their current movement and switch movements when they meet or collide.

Perspective switching

MultiviewC engine contains eight fixed location for camera acquisition where camera 8 is not involved in filming and acquisition during dataset acquisition. We have the freedom to switch the viewpoint to capture the specific image we need.

Multi-player simultaneous collection

To enable multi-view capture, the engine offers LAN online capability, allowing 50 players to capture online at the same time. The server will synchronise all players' collection operations and the data collected by each will be saved to the local machine.

Data collection and video recording

The engine features real-time video recording, data capture and more. The frequency of data acquisition can be customized. The data captured includes information on the rotation, dimension, position, action etc. of the target and the data is stored in the txt file.

--- Data collection ---

--- Video recording ---

Download

Download MultiviewC Engine from BaiduDrive(pwd:6666) or GoogleDrive. Notice: MultiviewC engine only supports windows system.

Toolkits for MultiviewC dataset

Overview

The MultiviewC dataset mainly contributes to multiview cattle action recognition, 3D objection detection and tracking. We build a novel synthetic dataset MultiviewC through UE4 based on real cattle video dataset which is offered by CISRO. The format of our data set has been adjusted on the basis of MultiviewX for set-up, annotation and files structure.

The MultiviewC dataset is generated on a 37.5 meter by 37.5 meter square field. It contains 7 cameras monitoring cattle activities. The images in MultiviewC are of high resolution, 1280x720 and synthetic animals in our dataset are highly realistic.

The simulation dataset not only provide 3D and 2D object detection annotation, but also simulated the common activities of cattle on the farm, including walking, running, eating, idle and sleeping. To fully exploit the complementary, the intrinsic and extrinsic parameters of each camera are also provided in MultiviewC dataset.

Project structure

File	Description
`data.py`	Methods for dataset visualisation and parsing labels, calibration information and images
`./utils/utils.py`	Methods for projecting 3D bounding box to 2D image and how to generate 3D bbox in our simulation dataset coordinate

Dataset

Download the data (images, annotation, calibrations) from BaiduDrive(pwd:6666) or GoogleDrive. We have updated a new annotation which has labeled the visibility of each cow at each frame.

The dataset folder structure is as following:

MultiviewC
├── annotations
|   ├── 0000.json
|   ├── 0001.json
|       ...
|       
├── calibrations
|   ├── Camera1
|   |   └── parameters.json
|   ├── Camera2
|   |    ...
|   └── Camera7
|        
└── images
    ├── C1
    |   ├── 0000.png
    |   ├── 0001.png
    |       ...
    ├── C2
    |   ...
    └── C7

Toolkits for MultiviewC dataset.

The repo includes the toolkits and utilities for building MultiviewC dataset.

How to's

download (from Baidu Drive Extraction Code: 6666 or GoogleDrive and copy the annotations, images and calibrations folder into this repo.
run the following command.

python data.py

done.

Jiahao-Ma / MultiviewC