Fast Hybrid Relocation in Large Scale Metric-Topologic-Semantic Map
read more
Citations
MapBased Navigation for a Mobile Robot with Ominidirectional Image Sensor COPIS
Recent trends in social aware robot navigation
Semantic Modeling of Places using Objects
Semantic loop closure detection based on graph matching in multi-objects scenes
Semantic representation for navigation in large-scale environments
References
Distinctive Image Features from Scale-Invariant Keypoints
Distinctive Image Features from Scale-Invariant Keypoints
Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope
Extremely randomized trees
Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials
Related Papers (5)
Frequently Asked Questions (14)
Q2. What are the future works in "Fast hybrid relocation in large scale metric-topologic-semantic map" ?
There are several ways to extend this work. Pixel-wise temporal consistency has been shown to improve labelling quality and higher level temporal consistency can be interesting to investigate.
Q3. What is the simplest way to enforce temporal consistency?
To enforce temporal consistency the authors accumulate Random Forest predictions from neighbours of the current view so that the unary potential ψc(yi) takes the form :ψu(yi) = α ∑ n∈N ψn(yi)where N is the neighbourhood of sphere i and α is a normalization factor.
Q4. What is the way to deal with dynamic environments?
Relocation under constraint: Navigation-oriented maps should provide an efficient way to deal with dynamic environments for lifelong mapping.
Q5. What is the way to capture the spatial structure of the scene?
To capture global structure of the scene a common solution is to embed first stage prediction results into a probabilistic graphical model [5].
Q6. What is the weighting strategy used between visual words?
The weighting strategy adopted between visual words is the term frequency-inverse document frequency tf-idf and the scoring type is L1-Norm (for details about parameters see [14]).
Q7. What is the Gibbs energy of a labelling?
In the fully connected pairwise CRF model the Gibbs energy [22] of a labelling y is:E(y) = ∑ i ψu(yi)+∑ i< j ψc(yi,y j)where ψc(yi) denotes φ(yi|X), ψu is the unary potential and ψc the pairwise potential.
Q8. Why is it called multivalent graph matching problem?
Due to change in viewpoint that can possibly fuse several objects, comparing those graphs formulates as multivalent graph matching problem.
Q9. What is the classification of a random forest?
A Random Forest is a set of T Decision Trees that achieves good classification rate by averaging prediction over all leaves L of all trees:
Q10. What is the dataset used for tests?
The dataset used for tests is the INRIA dataset presented in section III-C. CamVid is not used in this section because the dataset is too small with only 101 labelled images for sequence 06R0 and 124 for sequence 01TP.A.
Q11. What is the algorithm for detecting a pixel?
Their algorithm reaches near state-of-the-art performances for global per-pixel accuracy and outperforms [21] for average per-class accuracy.
Q12. Why is the image noise in the INRIA dataset so high?
It is due to the stitching algorithm used to build each view from three images that change locally the light intensity (please consult the video attachment of the paper).
Q13. What are the possible values for the map?
The possible values are: 1 = left, 2 = top left, 3 = top, 4 = top right, 5 = right, 6 = bottom right, 7 = bottom, 8 = bottom left.
Q14. What is the difference between the two graphs?
It is extremely lightweight: the size of the map with all full size images is 53Gbytes while semantic graphs representation needs only 18.5MBytes, whichcorrespond to a compression ratio around 3000.B. Interpretation Tree AlgorithmInterpretation tree is an efficient algorithm that uses relationships between nodes of two graphs to speed up the matching process.