Introduction

This is the website for our paper " Extrinsic Self-calibration of the Surround-view System: A Weakly Supervised Approach, "

Tongji Surround-view Dataset

dataset.zip(password: cblf)

We collected our own dataset by an electric car equipped with four cameras mounted around on campus. Our surround-view dataset mainly consists of two parts, calibration site images and natural ones. The calibration site images were collected over the calibration site to provide weak supervision information, while the natural ones, which act as training and testing sets, were taken from natural scenes. The original resolutions of all collected images are 1280×1080. Typical image samples contained in this dataset are shown below.

This original dataset contains 19,078 groups of fisheye images captured synchronously by SVS. In our implementation, the extrinsics' settings are augmented into nine different types, and the final dataset contains about 170,000 groups of images. 10% of the data were used for testing ( about 17,000 groups), and the rest forms the training set ( about 155,000 groups). In order to impove the performance of WESNet under different environmental conditions, images are grouped into 6 categories:

Source Codes

WESNetCode

Note: all these codes related to WESNet are implemented by Pytorch and we run them on Nvidia Titan Xp.

Demo Videos

The following are two demo videos demonstrating the performance of our WESNet in extrinsics calibration for synthesizing surround-view images and correction, respectively.

Last update: Aug. 16, 2021