Fitting Parameters for Procedural Plant Generation

Abstract

We propose a novel method to obtain a 3D model of a tree based on a single input image by fitting parameters for some procedural plant generator. Unlike other methods, our approach can work with any plant generator, treating it as a black-box function. It is also possible to specify the desired characteristics of the plant, such as the geometric complexity of the model or its size. We propose a similarity function between the given image and generated model, that better catches the significant differences between tree shapes. To find the appropriate parameter set, we use a specific variant of a genetic algorithm designed for this purpose to maximize similarity function. This approach can greatly simplify the artist's work. We demonstrate the results of our algorithm with several procedural generators, from a very simple to a fairly advanced one.

Single-View 3D Reconstruction via Differentiable Rendering and Inverse Procedural Modeling

“…This work generalizes and extends the results presented at the conferences [8,9]. The novelty of this work is that we propose a general approach for both differentiable and non-differentiable procedural generators as well as an intermediate case when only several parameters of procedural generator are differentiable (for example, the case of buildings).…”

“…This strategy was adapted in [8,47]. The work [47] analyzes an input tree image and infers a parametric L-system that represents it.…”

“…The work [47] analyzes an input tree image and infers a parametric L-system that represents it. In [8], the authors use more complex stochastic procedural models that simulate the process of vegetation growth, so the selection of procedural generator parameters is carried out using genetic algorithms. Both methods take a single tree image and find the optimal generator parameters by minimizing the loss function between a given reference and the model image created by the generator.…”

“…Even the simple procedural models demonstrated in the previous section prove to be extremely difficult to optimize functions for due to two factors: the large number of nonlinear internal dependencies within the generator and heterogeneous input parameters, some of which are integers with a limited set of possible values. Our first attempt at 3D reconstruction using procedural generators [8] relied on a specific implementation of a genetic algorithm [62] to find solutions without gradient calculation, but only for a limited set of problems. The ability to calculate the gradient of the loss function expands the range of available optimization methods.…”

Differentiable Procedural Models for Single-view 3D Mesh Reconstruction

“…Многие генераторы процедурных объектов работают как функции, которые преобразуют вектор числовых параметров в 3D-модель определенного класса, например, деревья или здания. Есть несколько работ, посвященных обратной процедурной генерации, которая включает в себя оценку входных параметров для генератора с учетом входной модели или изображения для создания моделей, аналогичных входным [28] [29] [30]. Однако, поскольку нетривиальные процедурные генераторы являются необратимыми функциями, существующие работы обладают ограниченной способностью к обобщению и сосредоточены на очень специфических процедурных моделях.…”