help

[xiangfengzhuo]

can you provide some references to this code? I'm so interested in this code

[midjji]

Huh? What are you looking for?

It's really just a single formulae, the code is primarily an evaluation toolkit.

On Thu, 14 Jun 2018, 20:14 xiangfengzhuo, notifications@github.com wrote:

can you provide some references to this code? I'm so interested in this code

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/midjji/ssba2016-scale-estimation/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AFMMJWuTwjDq87fX3OTAK6rUKuagap1aks5t8yaPgaJpZM4Uo9SR .

[midjji]

Ill add the pub to archivex, tonight, but if you are at cvpr we can meet and I can show you

[xiangfengzhuo]

Thanks for your reply，i just want to know how to estmate the monocular visual odomery's absolute scale in your code.I have been doing research in this area recently, but there is no progress.

在 2018-06-19 00:14:25，"midjji" notifications@github.com 写道： Huh? What are you looking for?

It's really just a single formulae, the code is primarily an evaluation toolkit.

On Thu, 14 Jun 2018, 20:14 xiangfengzhuo, notifications@github.com wrote:

can you provide some references to this code? I'm so interested in this code

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/midjji/ssba2016-scale-estimation/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AFMMJWuTwjDq87fX3OTAK6rUKuagap1aks5t8yaPgaJpZM4Uo9SR .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[midjji]

Hi, Ah unfortunately there is a misunderstanding,

The scale estimation in:

@inproceedings{persson2016metricscale, title={Metric scale with non-overlapping cameras, the 5+1 point method}, author={Persson, Mikael and Robinson, Andreas}, booktitle={Swedish Symposium on Image Analysis}, year={2016} }

also appended

Is about the case when you have cameras which are rigidly connected and take images at the same time, but unlike eg stereo their fields of view do not overlap.

The method is about a minimal case solver which is then refined by bundleadjustment.

From a pure geometric standpoint, it is strictly impossible to determine the true scale of a reconstruction from a single camera. T. he scale simply cannot be inferred from the data available.

But if we assume we have additional knowledge about what is beeing observed it is possible.

The simplest case is if say a chessboard pattern of known dimensions is observed and identified. In this case we can propagate this scale information throughout the posegraph, setting a true metric scale for the entire reconstruction. Such a solution drifts however. In practice the simplest thing to do is to use knowledge about some feature or object which frequently occurs and use that. Back when i worked with mono I used the license plates which in driving sequences are both everywhere and of a strictly regulated size. Another common approach is to use known camera placement relative to the road.

There are quite a few networks which output a depth map directly from a single image. This is a generalization of the approach above which combines object detection, known object size, common rigid 3D surfaces and works by training on a large dataset of similar data. In general they look like they almost work, but in practice they are terrible. Essentially its much better to train on a million annotated licence plates than use a few thousand images where most of the time the depth cannot be inferred with accuracy at all. Since depth networks can be trained in an unsupervised manner this might change, but its not quite there yet.

In the end the best solution depends on your application,

what are you working on?

//Mikael

2018-06-20 8:59 GMT+02:00 xiangfengzhuo notifications@github.com:

Thanks for your reply，i just want to know how to estmate the monocular visual odomery's absolute scale in your code.I have been doing research in this area recently, but there is no progress.

在 2018-06-19 00:14:25，"midjji" notifications@github.com 写道： Huh? What are you looking for?

It's really just a single formulae, the code is primarily an evaluation toolkit.

On Thu, 14 Jun 2018, 20:14 xiangfengzhuo, notifications@github.com wrote:

can you provide some references to this code? I'm so interested in this code

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/midjji/ssba2016-scale-estimation/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/ AFMMJWuTwjDq87fX3OTAK6rUKuagap1aks5t8yaPgaJpZM4Uo9SR .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/midjji/ssba2016-scale-estimation/issues/1#issuecomment-398645148, or mute the thread https://github.com/notifications/unsubscribe-auth/AFMMJSw7-nhxdxu3STkL3VQtM40xutIGks5t-fLTgaJpZM4Uo9SR .