Oil well reconstruction from monocular RGB videos

Reconstructing the oil well from pure images holds significant importance for improving and repairing oil wells. This process can offer valuable insights into the structural integrity of the well, allowing for better assessment and maintenance. The complexity arises from the need to accurately interpret visual data, such as identifying hidden or non-visible structural features and dealing with issues such as noise, low resolution, and distortions in the images. Therefore, we are devoted to offering fresh methods and creating high-quality 3D oil well geometric models from monocular RGB videos. To generate seamless reconstruction, we represent the oil well as the signed distance function field and conduct volume rendering on the signed distance function field to render the image. By minimizing the distance between rendered images and real images, we train the signed distance function field for acquiring the oil well geometric model. However, performing reconstruction purely from images is extremely challenging, and with only image supervision, the reconstructed geometric model is over smooth. To generate the detailed 3D geometry model, we propose to introduce dense point clouds created by multi-view stereo methods to supervise the training of the neural implicit signed distance function field. Our framework’s effectiveness and strength are clearly shown in the experimental results.