Evaluation

Benchmark Novel view synthesize: We utilize the same quantitative metrics on the test dataset as previous novel view synthesis works to evaluate the rendered 2D image quality: PSNR, SSIM, and LPIPS.

Benchmark 3d reconstruction: We use Chamfer distance to evaluate the quality of the reconstructed mesh. The groundtruth mesh retains faces that are similar to the 3D points visible from all viewpoints. You can click here to download them. If you used our download script, the groundtruth folder should already be included.

Before evaluating the estimated mesh, a preprocessing step isolates its visible part. Here, we follow the approach of NeuS. You can use our script for preprocessing. Compared to the script provided by NeuS, we have removed all faces below the lowest plane using groundtruth mesh. After isolating the visible part of the estimated mesh, a detailed quantitative analysis is conducted. This involves the uniform sampling of one million points from both the ground truth and the estimated meshes, followed by computing the Chamfer distance between these two point sets to quantify discrepancies.(see our repository for more information.)