Home > Free Essays > Tech & Engineering > Other Technology > Occupancy Grid for 3D Object Reconstruction

Occupancy Grid for 3D Object Reconstruction Research Paper

Exclusively available on IvyPanda Available only on IvyPanda
Updated: Sep 9th, 2021


The work is claimed to present an approach to construct photorealistic 3D models of real-world objects and to elaborate the methodology of system recognition of the movements. The system requirements are simple; therefore, the method is attractive for many applications. The computation of the visual hull is easy and fast and offers a good estimation of the object.

Object recognition and object reconstructions are comparatively new spheres in computer modeling and 3D modeling in particular. Object recognition is the sub-sphere of computer vision whose aim is to distinguish objects from image data and, often, to evaluate the positions and orientations of the defined objects in the 3D world. The images to be analyzed may be 2D gray-scale or color images or 3D range data images. Applications are many and comprise industrial machine vision, remedial image analysis, and content-grounded image recovery. Object reconstruction refers to the structuring of 3D object models from image or range data. Reconstruction of 3D surroundings from sensor-achieved information is a computer vision problem with important application to the area of virtual reality. The techniques for acquiring, registering, and fitting 3D data have been carefully discovered in the last few years. Prior work has been limited to particular objects and has merely manufactured a surface portrayal of the object as a whole. This paper addresses the task of recognizing entire 3D environments and objects, which necessitates scene fragmentation and image understanding systems. This is a knowledge-driven advance that endeavors to realize the physical structure of the surroundings and the particular objects in the environment through models that identify the physical properties and restraints of a particular field.


To make the model of the glove, and define the task for the computer on each finger movement, it will be necessary to make camera shots for every obvious movement (significant point, that the same camera or any other sensor was used by a computer). To simplify the task, it is necessary to mention, that the shots may not be done from all possible angles, but at least up to 20 grades from the front of the “glove”, as the operator will stand with one’s face to the sensor, but not with one’s back. The sensor should be able to recognize the movements, and to “teach” it the shots of the movements should be done (the 2D models, which represent series of shots for every movement)

Volumetric models, such as voxel-based ones, or using level-sets, are grounded on a discretization of 3D space and they aim to define the full and the unfilled cells. These processes can use a large number of images taken from randomly placed points of view. Any shape can be characterized and the visibility problem is handled in a deterministic geometric manner. However, the preliminary discretization limits the resolution of the objects. The only way of escalating the resolution is to increase the size of the voxel network. On the other hand, mesh depictions can, in theory, adapt their resolution to best renovate detailed shapes, but have difficulties dealing with self-intersections and topological modifications during the investigation.

Depth maps have been generally researched for two angles with a small baseline. The tiny baseline makes it unfeasible to get precise results and these techniques are forced to use tough priors that usually initiate front-parallel bias. The outlines of these techniques are not accurate permanent depth maps but piece-wise planar surfaces. Recently depth map reconstruction from multiple wide-baseline images has been developed with impressive results. The wide-baseline arrangement allows amazingly accurate results without the discretization and topological difficulties of other methods. These good properties of the depth map depiction stimulate us to use it. Nevertheless, a single depth map is frequently not enough to represent the whole sight: only the parts viewed in an orientation view are sculpted. A depth map for every input picture is necessary to guarantee that every input pixel is utilized and modeled. This is probably the model best adjusted to the resolution of the input and is the model regarded the work. Alternatively to computing each depth map separately and including them in a post-processing step, such model aims to calculate all the depth maps at the same time which authorizes an efficient geometric visibility/ occlusion analysis and guarantees that the outlined depth maps will be consistent.


In-depth map recovery was invented as a maximum a posteriori (MAP) issue using the framework proposed for the novel view combination problem showing that the two issues are the same. Here we adopt this frame and adapt it to the case of multiple orientation views. The main contributions of this paper to this framework are: First, a reflection on and modification of the probability formula. The second is geometric visibility prior. We use the current depth maps judgment to determine the prior visibility of the reproduction points. And in conclusion, multiple depth maps before that flats and combines the depth maps while preserving discontinuities.

The goal is to find a 3D representation of a scene, from a known set of images with full calibration data, i.e. known intrinsic and extrinsic factors. The model which is used to characterize the scene comprises a set of colored depth maps. For every pixel in the key images, the aim is to infer the depth and color of the 3D point that this pixel is seeing.

The main problem which is faced is the issue of a Bayesian MAP search. Input images I am regarded as a noisy measurement of the model Θ. The researched model is defined as the one that maximizes the posterior probability p(ΘǀI) ∝ p(ΘǀI)p(Θ).

Having all the visibility variables identified, the next step in a Bayesian modeling assignment is to choose the field of the decomposition of their joint probability. The disintegration will define the statistical dependencies among the variables that our model is considering. For completeness, we add to the previously defined variables, a variable, that signifies the set of all the parameters that will be used in the current method. The joint probability of all the variables is then p(I, V, I*;D, τ ) and the offered decomposition.

  1. p(τ) is the previous probability of the factors. A consistent one in this work and ignore this term is assumed.
  2. p(I| τ ) is before the colors of the depth maps. The so-called image-grounded priors were commenced to enforce the calculated images I* to look like usual pictures, which in practice was forced by looking like images of a catalog of instances.
  3. p(D| τ) is prior on depth maps. Its work is flat and incorporates diverse depth maps. Modeling this confidence can help when arranging with stable albedo surfaces where image and depth discontinuities are compared.
  4. p(V|D, τ) is the visibility prior. It is offered to regard visibility as reliant on D, to allow geometric reasoning on occlusions. In the E-step of the EM algorithm depicted below, this geometric visibility prior will be probabilistically assorted with photometric substantiation, giving an estimation of the visibility that is more robust to arithmetical occlusions than using a uniform prior.
  5. p(I|v, I*, D, τ ) is the probability of the input images. Meticulous interest is paid to this term because it is found that typical formulae are not reasonable for the wide-baseline case.


The proposed method was created in meticulous probabilistic frames expanding previous related works. The researches proved the pertinence of these additions. However, there are still some matters to solve to make the technique more utilizable. The probabilistic advance allows the parameters of the technique to be learned during the optimization. In effect, treating the options as random variables it is possible to either approximate their most possible value or marginalize them out. The current implementation requires setting three options. Although these parameters represent well-defined concepts it will be preferable that the algorithm automatically sets them.

The other problem of the method, like in any other incline tumble-based scheme, is the initialization. The pyramidal completion of the EM algorithm meets well in cases where the brawny discontinuities are imprisoned from previous small resolution levels. Nevertheless, without a good initialization, it appears likely that images such as the ones regarded in the EM algorithm does not attain the total optimum but only the local one. Fascinatingly, one of the best performing techniques in this area uses the same Bayesian method, but the optimization is done with the Loopy Belief proliferation algorithm. It is our attention to study the opportunity of applying this or other universal maximization methods to our subsequent likelihood description.


Gargallo, P. Sturm, P 2004 ‘Bayesian 3D Modeling from Images using Multiple Depth Maps’ INRIA Rh.one-Alpes, GRAVIR-CNRS, Montbonnot, France.

Gargallo, P. Sturm, P. Pujades, S 2007 ‘An Occupancy–Depth Generative Model of Multi-view Images’ INRIA Rhˆone-Alpes and Laboratoire Jean Kuntzmann, France.

Rocha, R. Dias, J. Carvalho, A. 2005 ‘Cooperative multi-robot systems: A study of vision-based 3-D mapping using information theory’ Institute of Systems and Robotics, Faculty of Sciences and Technology.

This research paper on Occupancy Grid for 3D Object Reconstruction was written and submitted by your fellow student. You are free to use it for research and reference purposes in order to write your own paper; however, you must cite it accordingly.
Removal Request
If you are the copyright owner of this paper and no longer wish to have your work published on IvyPanda.
Request the removal

Need a custom Research Paper sample written from scratch by
professional specifically for you?

801 certified writers online

Cite This paper
Select a referencing style:


IvyPanda. (2021, September 9). Occupancy Grid for 3D Object Reconstruction. https://ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/


IvyPanda. (2021, September 9). Occupancy Grid for 3D Object Reconstruction. Retrieved from https://ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/

Work Cited

"Occupancy Grid for 3D Object Reconstruction." IvyPanda, 9 Sept. 2021, ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/.

1. IvyPanda. "Occupancy Grid for 3D Object Reconstruction." September 9, 2021. https://ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/.


IvyPanda. "Occupancy Grid for 3D Object Reconstruction." September 9, 2021. https://ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/.


IvyPanda. 2021. "Occupancy Grid for 3D Object Reconstruction." September 9, 2021. https://ivypanda.com/essays/occupancy-grid-for-3d-object-reconstruction/.


IvyPanda. (2021) 'Occupancy Grid for 3D Object Reconstruction'. 9 September.

Powered by CiteTotal, free citation creator
More related papers