search

uzh-rpg / rpg_open_remode

This repository contains an implementation of REMODE (REgularized MOnocular Depth Estimation), as described in the paper.
http://rpg.ifi.uzh.ch/docs/ICRA14_Pizzoli.pdf
GNU General Public License v3.0
491 stars 187 forks source link

Noisy cloud and low convergence rate #27

Open Bardo91 opened 6 years ago

Bardo91 commented 6 years ago

Hi, I am setting up an UAV with a ZED camera and a Jetson tx1. I am testing REMODE with a custom ZED wrapper that publish SVO message to fit to the ros interface. Till now everything seems to be ok, but the cloud I am getting is very noisy, and the convergence rate never grows up more than ~21.6%. I uploaded a gif my results. I decreased the convergence threshold to 20%. Does anyone have any clue of what could be my error?

5a574bef39a0d496223361

Thanks in advance

Tsugikuni-Yoriichi commented 5 years ago

I have the same problem. The first depth image seems produce a pyramid structure. I am not sure how it's caused. I think the author has already write the denoise process in his code. Is it because that the denoise parameter should be revised according to our own dataset. I find that in "depthmap_denoiser.cu" there are parameters for denoising. Anyway, have you solved this problem? Even I removed the first depth image, the remainning part is not in very good condition as well.

Tsugikuni-Yoriichi commented 5 years ago

I just jumped the first part of video. It gives now better result than previous ones. But still a further smooth on depth map is required for noise I think. Yours seem to select the key frame so frequent. Maybe key frame selection speed should be in a reasonable range. It can be changed in config.cpp of svo with kfselect_mindist.