SimpleSFM icon indicating copy to clipboard operation
SimpleSFM copied to clipboard

Problem with point cloud results

Open hbzhang opened this issue 6 years ago • 7 comments

When I view the point cloud, it does not looks right.

It looks like this (I uploaded jpg here)

http://i64.tinypic.com/160ealc.jpg

All different directory seems return same point cloud data for the following directory:

TestSeriesMe TestSeriesPoster

hbzhang avatar Feb 08 '18 15:02 hbzhang

I use Matlab plot the ply file.

hbzhang avatar Feb 08 '18 15:02 hbzhang

I'm guessing this is just because the LM optimizer found a horrible local min in bundle adjustment. There are some fixes you can try, but you should know that I hacked this project together pretty quickly as an exercise in teaching myself about SfM. If you want a real SfM toolbox I would recommend one of the many larger, well-maintained libraries (e.g. bundler, openmvg, etc).

dfridovi avatar Feb 08 '18 17:02 dfridovi

Thank you for very quick reply. I indeed like your work and I have actually spent time follow this work. Thing is here, on your machine, can you run the demo images with good point cloud? I wonder why these demo images do not work on my machine.

I have not changed any your code besides I modify the Pts3D from dictionary to array, which has been reported previously by another developer.

Any hint?

Also, appreciate your recommendation of other packages.

hbzhang avatar Feb 08 '18 18:02 hbzhang

To be honest, it's been so long since I last touched this project that I don't remember how reliably it could create point clouds. I think I remember running one test where I took pictures of stuff on a wall and then looked to make sure all the points were co-planar in the point cloud (and they were). I guess the message here is that all the major components in a simple SfM pipeline are here, but they are definitely not reliable. So if you want to learn about SfM and mess around a bit on your own, this is a nice starting point; but if you want something that actually works well I'd suggest finding a well-maintained SfM project like the above.

dfridovi avatar Feb 08 '18 19:02 dfridovi

Thank you. It makes perfect sense!

Hongbo

On Thu, Feb 8, 2018 at 2:11 PM, David Fridovich-Keil < [email protected]> wrote:

To be honest, it's been so long since I last touched this project that I don't remember how reliably it could create point clouds. I think I remember running one test where I took pictures of stuff on a wall and then looked to make sure all the points were co-planar in the point cloud (and they were). I guess the message here is that all the major components in a simple SfM pipeline are here, but they are definitely not reliable. So if you want to learn about SfM and mess around a bit on your own, this is a nice starting point; but if you want something that actually works well I'd suggest finding a well-maintained SfM project like the above.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/dfridovi/SimpleSFM/issues/2#issuecomment-364217040, or mute the thread https://github.com/notifications/unsubscribe-auth/AAMj1ShpSuwu2Zk-OqdruBUgtC-u13Pdks5tS0b4gaJpZM4R-j1C .

hbzhang avatar Feb 08 '18 22:02 hbzhang

When I view the point cloud, it does not looks right.

It looks like this (I uploaded jpg here)

http://i64.tinypic.com/160ealc.jpg

All different directory seems return same point cloud data for the following directory:

TestSeriesMe TestSeriesPoster

My result is similar to you. The result is not right....... Did you find any solution or better sfm easy repo?

weihaosky avatar Oct 23 '18 02:10 weihaosky

Hi there. As I've mentioned above, I want to emphasize that this repository was intended as a learning exercise and not a working demo. If you want to use this repository to see all the fundamental building blocks, that's great, but please don't expect it to work all that well.... though if you do make your own changes and want to open a PR I'm certainly willing to look at it.

dfridovi avatar Oct 23 '18 03:10 dfridovi