Link to the web app.
Please note that it may take a few minutes for the web app to start up, in case it has not been run recently. That's because the web service is hosted through a free tier that puts idle services on standby.
ViDA Summary has been carried out by Jacopo Massa as part of the Scientific & Large Data Visualization course at the University of Pisa, under the supervision of teachers Daniela Giorgi and Massimiliano Corsini.
Analyzed models are a subset of the ViDA 3D dataset, in particular there are 6970 models described by some attributes, including:
- Name
- Categories
- #Likes
- #Views
Model analyzes were divided into two web pages, described in subsequent sections.
Bar Chart and Treemap show a subsantial imbalance in favor of only 3 categories:
- Characters & Creatures
- Architecture
- Cultural Heritage & History
With the scatter plot, correlation between any couple of models' numerical attributes can be observed.
Models were given in input to a
ResNet,
in particular a ResNet50
, that extracted 2048 features per model.
Starting from these data, a simplified version of the IsoMatch algorithm was applied, in order to obtain a compact grid representation and possibly find some correlation between the analyzed models.
First step of the algorithm consists in reducing the dimensionality
features (in this case dim = 2
to represent models on 2D
Cartesian graphs).
A first representation was obtained by applying a clustering algorithm (K-Means) to the "reduced" data. Clustering has confirmed what was observed in the first part, that is, that models can be grouped in the 3 most common macro-categories.
Finally, models were placed in a variable size grid (based on each cluster's cardinality). This representation provided a summarized view of the analyzed subset.