search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.23k stars 242 forks source link

2 Stage Training , RANSAC method Question #189

Closed sjg02122 closed 2 years ago

sjg02122 commented 2 years ago

Thank you very much for your work. This is very helpful for my research.

I have a question about how to solve the infinite depth problem using RANSAC (2-stage training).

  1. I wonder how much the performance improved when re-sampling was performed using RANSAC.

  2. I understand that the method of re-sampling is as follows.

2-1) Point Clouds are created using Depth Estimation. 2-2) The ground plane of Point Clouds is estimated using RANSAC. 2-3) Compare each pixel of the estimated Ground Plane and Depth map to filter things with significantly different values.

I understand that this process is carried out by designating threshold At this time, when the threshold point is 10, does it mean that the number of sensed pixels is 10?

VitorGuizilini-TRI commented 2 years ago

Hi, thank you for your interest in our work! This threshold means that we are evaluating all pixels to find the ground-plane, and removing frames if there are more than threshold pixels below the extracted ground plane. Let me know if that helps!