Creating Objects and Object Categories for Studying Perception and Perceptual Learning

This procedure aims to create objects and object categories for studying how we perceive and learn to perceive objects by sight and or touch. First, virtual morphogenesis or VM is used to simulate the processes of early embryonic development and create novel naturalistic virtual 3D objects called digital embryos. Then using virtual phy agenesis or VP object categories are created with precisely defined statistical properties based on the input digital embryo.

If desired principle components analysis can be used to create additional shape variations among the virtual objects created by virtual morphogenesis and virtual phy agenesis. The likelihood that a given object belongs to a given category can be precisely calculated using feature-based bayesian inference. If necessary, haptic prints of the resulting virtual objects can be generated using a 3D printer.

Each of these methods will be illustrated in greater detail later Compared to existing methods. These new approaches create naturalistic but precisely measurable shape variations that arise without the need for investigator imposed constraints. They offer novel tools in the fields of visual and haptic perception, perceptual learning, and machine vision, and have potential applications in the rehabilitation of many types of visual impairments through visual haptic cross metal training.

Interestingly, this method can also be applied to study the processes of morphogenesis and evolution themselves, And we first had the idea for this method when we were looking for ways to generate naturalistic, but precisely definable visual stimuli for studies and computational vision. Initially, individuals knew through this method might struggle with its math and programming intensive aspects, so this visual demonstration will illustrate how to implement and use this method properly. In the digital embryo workshop, specify a set of settings or genotype to generate a single embryo to generate multiple embryos.

Repeat this process multiple times to generate more complex shapes by virtual morphogenesis. Increase the number of growth cycles to specify the number of times the cells of the embryo will divide. The digital embryo workshop automatically saves each embryo as a OBJ file so you can later use the embryo with commercial 3D modeling toolkits.

Generate the visual stimuli by setting the various standard graphical parameters such as orientation, size, lighting, surface texture, and background to generate object categories. Create descendants of an ancestor object in a hierarchical manner. You may also smoothly vary the shape using morphing while preserving the one-to-one correspondence of vertices among the objects.

Interestingly, virtual objects other than digital embryos can also be used as inputs to virtual phy agenesis. Select objects within a given category to achieve a given distribution of features. For instance, if you want to create two categories that differ in size, selectively eliminate mid-sized objects to generate a bimodal distribution of object sizes.

Now, objectively measure the similarity between a given pair of categories using available phylogenetic methods such as co phonetic correlation. These calculations can be carried out using commonly available analytical toolkits such as MATLAB or R for any given pair of objects where each vertex of one object corresponds to exactly one vertex of the other object morphing is straightforward. Select the interpolation points and use linear morphing between the two objects to smoothly interpolate between the corresponding vertices.

First, determine the principle components as specific descriptors of a given set of objects. Principle components can be calculated using MATLAB or R average, the coordinates of each vertex across all N input objects to generate an average object, multiply any component P by the corresponding egen value lambda and a desired weight wj, and add to the average object to generate a new object. Aj continue to smoothly vary WJ to create smooth shape variations along a given principle component.

To create a multi-dimensional grid of shapes, use a set of weights for each of several principle components. Print out 3D objects using a 3D prototyper. If necessary, adjust the object's size and smooth the object's surface to optimize print.

An important task in visual processing is inferring the category to which a given observed object belongs. In part by using the information about known features of the object, digital embryos are useful. When studying this inferential process, let's assume that the categorization task is binary.

That is there are only two possible categories and that our task involves distinguishing category K from category L, let C be the category variable, C equals K or C equals L according to whether the observed image I belongs to category K or L respectively. Assuming that there is exactly one binary feature F, calculate a probability that the category is K given the information in the image. Similarly, for the probability that the category is L, pick the category with the higher probability.

For example, start with this informative fragment feature and a threshold value of 0.69. The task is to determine whether this feature is present in a given image like the rightmost image in road G three. First slide the template over all possible locations in the image compute at each location, the absolute value of normalized cross correlation between the template and the underlying sub image.

Then select the image location with the highest value. If this value is above the threshold, conclude that the feature is present, otherwise conclude that it is absent. Within the framework of feature based inference, we assume that all the information the observer extracts from the image is contained in the value of this feature.

Therefore, the task becomes that of determining the value of F in the given image computing probabilities for that value F, and selecting the category with the higher probability. This is the Bayesian framework for putting together all of the relevant probabilities. Note that the denominator in the two equations is the same, so restrict attention to the numerator.

Assume a flat prior that is both categories are apriori. Equally likely the task is now to compute likelihood the probability of a given feature value in an image of a given category C.For example, use the six images of category L as examples to compute the probability that the feature is present in an image of category L.First, take all the training images that belong to L for each image, determine whether the feature value is one that is the feature is present in the image, or zero that is the feature is absent. Then compute the fraction of images in which the feature value is one.

Therefore, the probability that the feature is present in an image of category L is 0.33 for accurate estimates, use at least 30 images per category. In a typical experiment, we would need to know the subject internal estimate of this probability. Note how using digital embryos makes this especially easy.

Since we have full control over the subject exposure to digital embryos, we can be sure that the subject internally computed value is consistent with our estimate and is not influenced by any uncontrolled and unknown previous experience. In a similar manner, compute probabilities of absence and presence of the image in categories K and L.Given these values, inference can be performed to identify the category label of this new image. First, determine whether the feature F is present in the image using the previous formulas determined for un-normal probabilities and the values just computed, calculate the probabilities of the presence in the image of categories, K and L.These data indicate that the image is from category K.Although with relatively low confidence, virtual morphogenesis offers a limitless supply of novel 3D shapes.

Here, digital embryos are generated by simulating key processes of biological embryogenesis. Each run starts with an icosahedron and generates a unique embryo. Based on the morphogen settings, the digital embryos can be graphically manipulated to create visual scenes of arbitrary complexity using any standard graphical toolkit.

For example, the same digital embryo can be textured differently and lit as desired. In addition, visual scenes of arbitrary complexity such as this scene with a digital embryo camouflaged against a similarly textured background can be created using a commercially available 3D modeling and rendering environment. The virtual phy agenesis algorithm emulates biological evolution.

The virtual phy agenesis algorithm emulates biological evolution. Novel objects and object categories emerge as heritable variations that accumulate selectively but accrue shape variations of their own as they develop. In this particular example, a single common ancestor, the icosahedron produces three generations of descendants.

The shape complexity increases from the icosahedron to generation G one because we allow the number of cells to increase, but the overall shape complexity stays the same from Generation G one onwards. This family tree is comparable in other respects, but it uses non embryo objects which were downloaded from virtual object vendors. Notice that the objects that share a common ancestor straightforwardly constitute a category.

Since no cell divisions were allowed in any generation, all shape variations result solely from the movement and or growth of the individual cells of the given object. In this scenario, morphing creates smooth variations in shape by interpolating between the corresponding vertices of the two designated objects. The far left and the far right.

Embryo principle components also create smooth variations in shape. This embryo represents the arithmetic average of 400 embryos. In this particular case, the first two principle components accounted for 73%and 19%of the shape information respectively.

Embryos were obtained by varying the weighted eigen values. These digital embryos can be rendered as virtual 3D objects and then printed as haptic objects using a standard commercially available 3D printer or prototyper for studying visual perception as inference, in particular as Bayesian inference. Digital embryos are an invaluable tool for creating novel categories with controlled parameters such as priors and likelihoods.

After watching this video, you should have a good understanding of how to create a set of digital embryos suitable for your particular experiment. Individual objects or entire categories with different degrees of variability and complexity can all be created easily. The resulting images can be used for experiments in object recognition, categorization, category learning, and many others.