Hey thx again for uploading the SLAM-Course. Whats the date of the next Video? ^^
@007billgates4 жыл бұрын
Thank you for the lecture. For better understanding implementation wise, FastSLAM 2 algorithm should be in some slide. Similar to what was there for FastSLAM 1.0
@robodoctor6 жыл бұрын
well explained !
@shrinivasiyengar57992 жыл бұрын
If high motion noise was a problem in FastSLAM applied to grid-based maps and mapping with known poses fails, then it should also have failed for FastSLAM 1.0 for landmark-based SLAM, right?
@truonggiangnguye3230 Жыл бұрын
I don't understand samples around position maximum likelihood. How to implement it?
@shrinivasiyengar57992 жыл бұрын
I am sure the answer to the question is yes, but I still want to shoot my shot: When we are not resampling so frequently then after the particles get their weights readjusted based on the measurement correction, we let the particle go ahead with the next iteration of proposal distribution sampling. But before we go to the next iteration of proposal distribution sampling do we normalize all the weights?
@shrinivasiyengar5799
2 жыл бұрын
Giving it a second thought, naturally they must have been normalized after weight/measurement correction. That is the only way \eta_{eff} stays between the number 1 and the number of samples.
@dmitrygavrilenko5318 жыл бұрын
Why do we need weights for FastSLAM 2 particles if we generate particles, already taking into account measurements?
@CyrillStachniss
8 жыл бұрын
+Dmitry Gavrilenko Because weight = target / proposal and if we change the proposal, the weight will be different as well.
@dmitrygavrilenko531
8 жыл бұрын
+Cyrill Stachniss Thank you for your response! My question was more about understanding the general idea of the method rather than about math formulae. If the proposal distribution already incorporates measurements, what is the purpose of the target distribution at all?
@Pages_Perfected2 жыл бұрын
hmm, is there a ready program to do the mapping side?
@rafaelhsouza7 жыл бұрын
At around 40:00, why are the samples taken from tau and not from p(xt|...)?
@abhisekpanda5105
7 жыл бұрын
I think the Gaussian distribution obtained is approximately equal to PROPOSAL distribution and NOT tau. Hence sampling should be taken from Normal distribution( != tau).. As for the lecture, he may have said it incorrectly, but you can read his paper(available on the course website) where it is clearly explained...
@shrinivasiyengar5799
2 жыл бұрын
@@abhisekpanda5105 I think he is taking samples from \tau(.), because he mentions on multiple occasions that \tau(.) needs to be evaluated at these sample locations. Then he goes on to say that the computation drawing the samples and having \tau evaluated at them helps us in approximating the integral in the denominator as well. Evaluating these two things gives us finally the proposal distribution.
@shrinivasiyengar5799
2 жыл бұрын
@@abhisekpanda5105 I have not read the paper yet, and am only basing this on the lecture. Maybe the paper says something else, but this is what i understood till now.
Пікірлер: 15
Hey thx again for uploading the SLAM-Course. Whats the date of the next Video? ^^
Thank you for the lecture. For better understanding implementation wise, FastSLAM 2 algorithm should be in some slide. Similar to what was there for FastSLAM 1.0
well explained !
If high motion noise was a problem in FastSLAM applied to grid-based maps and mapping with known poses fails, then it should also have failed for FastSLAM 1.0 for landmark-based SLAM, right?
I don't understand samples around position maximum likelihood. How to implement it?
I am sure the answer to the question is yes, but I still want to shoot my shot: When we are not resampling so frequently then after the particles get their weights readjusted based on the measurement correction, we let the particle go ahead with the next iteration of proposal distribution sampling. But before we go to the next iteration of proposal distribution sampling do we normalize all the weights?
@shrinivasiyengar5799
2 жыл бұрын
Giving it a second thought, naturally they must have been normalized after weight/measurement correction. That is the only way \eta_{eff} stays between the number 1 and the number of samples.
Why do we need weights for FastSLAM 2 particles if we generate particles, already taking into account measurements?
@CyrillStachniss
8 жыл бұрын
+Dmitry Gavrilenko Because weight = target / proposal and if we change the proposal, the weight will be different as well.
@dmitrygavrilenko531
8 жыл бұрын
+Cyrill Stachniss Thank you for your response! My question was more about understanding the general idea of the method rather than about math formulae. If the proposal distribution already incorporates measurements, what is the purpose of the target distribution at all?
hmm, is there a ready program to do the mapping side?
At around 40:00, why are the samples taken from tau and not from p(xt|...)?
@abhisekpanda5105
7 жыл бұрын
I think the Gaussian distribution obtained is approximately equal to PROPOSAL distribution and NOT tau. Hence sampling should be taken from Normal distribution( != tau).. As for the lecture, he may have said it incorrectly, but you can read his paper(available on the course website) where it is clearly explained...
@shrinivasiyengar5799
2 жыл бұрын
@@abhisekpanda5105 I think he is taking samples from \tau(.), because he mentions on multiple occasions that \tau(.) needs to be evaluated at these sample locations. Then he goes on to say that the computation drawing the samples and having \tau evaluated at them helps us in approximating the integral in the denominator as well. Evaluating these two things gives us finally the proposal distribution.
@shrinivasiyengar5799
2 жыл бұрын
@@abhisekpanda5105 I have not read the paper yet, and am only basing this on the lecture. Maybe the paper says something else, but this is what i understood till now.