QSM-FaNNI stands for Quantitative Structure Models - Foliage and Needles Naïve Insertion, and it is an algorithm and a Matlab implementation for generating leaf covers for structural tree models. Leaf shape, size, orientation and leaf area distribution on the tree structure is highly customizable. The resulting leaves are non-intersecting clones of the selected leaf basis geometry that are associated with a parenting geometric primitives of the structure model. The source-code of the Matlab implementation is freely available of the research group's GitHub repository.

The video showcases the results of the algorithm on three reconstructed oak tree models, as well as, some of the customization possibilities, namely the leaf area density distribution (LADD) and leaf basis geometry shapes. The first example LADD is featured on the corresponding research article.

A quantitative structure model (QSM) contains the geometric and topological structure of a reconstructed tree. As such, QSMs enable computation of detailed tree properties that have been laborious or impossible to measure before. The computed tree properties can be used as classification features for tree species recognition.

The first half of this video illustrates how we define the 15 classification features our research group has used for a species recognition study. An example QSM is used to visualize the relevant tree parts and key steps in the feature computations.

The second half shows how the feature values of over a thousand Finnish trees of three different species, Silver birch, Scots pine and Norway Spruce, are distributed, and how well the species separate in the defined feature dimensions. One example QSM of each tree species shown on the right-hand-side with the only the tree parts visible that are related to the current feature.

The contents of this video link directly to the paper titled Automatic tree species recognition with quantitative structure models.

Forward modeling and inversion in a volumetric domain: Brain activity recovered based on magnetoencephalography data with a divergence conforming source model and a hierarchical Bayesian approach.

We develop volumetric forward and inversion methodology for various technologically advanced applications including (1) the detection and stimulation of low-frequency (quasi-static) biopotential fields and (2) the high-frequency (waveform) radio/microwave/acoustic imaging systems of astro-/geophysics, biomedicine and non-destructive material testing. Characteristic to (1) is, among other things, that the complexity of the biological tissue structures, such as the human brain, can be a major challenge that necessitates careful mathematical modeling. In (2), the sparsity of the measurements and the large computational effort related to simulating and inverting full waveform data can set additional requirements for the imaging process.

Summary of how Quantitative Structure Models are reconstructed, and how they are being utilized. The process starts with the laser scanning of a forest. The produced point cloud contains millions of points, and numerous trees. The point cloud is segmented automatically into trees and then into branches. Each branch is reconstructed by fitting several cylinders to the data. The branching topology is stored together with the geometric structure to a single tree model, which is called a Quantitative Structure Model (QSM).

QSMs can be used in many applications, as they allow easy access to tree properties, such as, volume, area, branch count, crown shape, taper curve, and size distribution. These properties can be used by forest scientists, forest industry, and forest owners to, e.g., evaluate the current or future state of a forest, or its value. The model properties can also be used for automatic species recognition. Furthermore, QSMs can be visualized in various ways to produce realistic virtual representations of locations, such as, national or city parks, for the travel industry. The models can be textured and augmented with leaves to achieve either a realistic or a fantasy look, to suite the needs of game developers.

Visualization on how a point cloud is created by a terrestrial laser scanner, and how a reconstructed quantitative structure model can be visualized in various ways. The visualization can be a cylinder model as shown in previous videos, but it can also be converted to a more continuous Bézier surface. The video also shows how the tree model can be augmented with non-intersecting leaves by sampling a certain distribution based on the branching structure. Both the branches and the leaves can be textured either for realism or something totally different.

This is the Finnish version of the Cylinder reconstruction video listed below.

Kattava puurekonstruktio sylinterisovitukseen perustuen. Laserkeilaamalla kerätty pistepilvi segmentoidaan oksiin ja edelleen osasegmentteihin, jotka rekonstruoidaan yksitellen joukkona sylintereitä. Jokainen sylinteri sovitetaan pistepilven osasegmenttiin pienimmän neliösumman menetelmällä. Puun oksautumisrakenne eli topologia siirtyy segmentoinnista osaksi kattavaa sylinterimallia, jota kutsumme termillä quantitative structure model (QSM).

Segmenttitasolla prosessi alkaa segmentin juuresta, johon lisätään "pistekerroksia", kunnes osasegmentin haluttu pituus--leveys-suhde saavutetaan. Alkuarvot pienimmän neliösumman iteratiiviselle sovitukselle löydetään hyödyntämällä geometrista analyysia osasegmentin muodostamaan pistepilveen. Jos sovitus ei suppene tai jos pistepilvidata on liian harvaa, voi mallin sylinterien väliin jäädä liian suuria välejä. Välit voidaan paikata jälkikäsittelyvaiheessa, jossa analysoidaan tyhjän tilan määrää peräkkäisten sylinterien sekä vanhempi--lapsisylinterien välissä.

Kun yksittäisen segmentin rekonstruktio on valmis, käytetään heuristiikkaa estämään esimerkiksi liian suuret erot peräkkäisten sylinterien säteissä.

Comprehensive tree reconstruction based on cylinder fitting. A segmented point cloud received from terrestrial laser scanning is reconstructed one subsegment at a time as a collection of cylinders. Each cylinder is fitted in the least-squares sense to the point cloud data corresponding to a subsegment. The branching topology from the segmentation is transferred and stored in the resulting cylinder model which we call a quantitative structure model (QSM).

For each segment the procedure begins from the segment base by selecting layers of neighboring points using the neighbor relation defined by the surface patches. These layers are appended to the current subsegment until a selected length/radius ratio is achieved. Geometrical analysis is used to find good initial values for the iterative least-squares cylinder fitting. If fitting does not converge, or point data is insufficient at some point, gaps can occur in the model. As a post-process some of the gaps are filled by analyzing the empty space between consecutive cylinders in the topology and also between child and parent cylinders.

When a segment is completed heuristics are used to further tune the properties of the cylinders to prevent, e.g., large differences between the radii of consecutive cylinders.

This is the Finnish version of the 3D Forest Information video listed below.

Animaatiossa näytetään, kuinka metsäpalsta voidaan rekonstruoida automaattisesti maanpäälisestä laserkeilausdatasta. Yksittäiset puut tunnistetaan ja mallinnetaan sylintereillä. Sylinterimalleista voidaan laskea puiden ja metsän kannalta hyödyllistä tietoa.

Animaatiossa käytetään aitoa laserkeilausdataa, jonka on kerännyt Eric Casella, joka työskentelee Forest Research -instituutissa Isossa-Britanniassa. Laserkeilaus suoritettiin huhtikuussa 2012. QTVR-formaatin visualisoinnit pistepilvistä ovat saatavilla ResearchGate-palvelussa.

This video shows how a forest plot can be automatically reconstructed from terrestrial laser scanning (TLS) data. Each of the trees are identified in the data and reconstructed as cylinder models. The resulting models can be used to derive valuable tree and forest information.

The animation is based on real TLS data collected by Eric Casella working at Forest Research (UK) in April 2012. QTVR visualization of the point clouds are available at ResearchGate.

Automatic segmentation of a point cloud presenting a single tree captured by terrestrial laser scanning. The approach uses the neighbor relation of the surface patches whose center points are presented as spheres in this animation. The process starts from the tree base and advances until all surface patches are processed.

At every step, a cut set is defined for the current segment. The cut set is basically a layer of surface patches that are either appended to the current segment or marked as a base of a new segment. Whenever the cut set becomes disconnected, each of its components is inspected in various ways to determine whether it belongs to the current segment or not. A component that is not part of the segment is marked as a new base. A segment is completed when it cannot be extended into any unprocessed direction. Upon completion a new base is selected from the process queue.

Classification method of a point cloud presenting a tree. By using surface patches and their geometric properties and also the well-defined neighbor-relation, it is possible to automatically classify surface patches into several categories: ground, trunk and branches. The animation uses real terrestrial laser scanning data from a single tree.

The trunk is located by using filtering based on geometric properties. Then the trunk base is determined by fitting a cylinder to the trunk and by sliding it downwards until its radius changes drastically. When the base is defined, ground points are excluded by selecting them through extension starting from the base and not going into the trunk. Points that are not part of the trunk nor the ground are initially labelled as branch points.

Introduction on how geometric characteristics are evaluated for a point cloud, which can be, e.g., a surface patch. First part shows how principal components depend on the data. When principal components are computed for every individual surface patch, they can be used to find dimensionality and direction estimates for them. The end part of the animation shows how the principal components of neighboring surface patches can be used for a more accurate direction estimate by finding a direction that perpendicular to the plane spanned by the smallest principal components.

Animation that shows how a point cloud can be partitioned into subsets called surface patches. These surface patches can be used to describe the local geometric properties of the point cloud. As each patch knows its neighbors, the patches can also be used to move along the point cloud through set expansion.