First clustering via pixplot shows some clusters by colors and strong features such as visible grids or patterns. Not much visible in terms of details.
Filtering by year reveals that early games are in the global center or in the center of clusters. With 1982 comes an explosion of interfaces that are dispersed all over the edge-regions, which could be an interesting finding. As in, 1982 starting the time of experimentation. I need to check how this can be developed further.
Following Janna Omena’s lead on reflecting our digital research methods (see Technicity), I investigated the base approach of pix-plot. It basically uses Tensorflow’s Inception v3 model for feature extraction and UMAP for the reduction of dimensionality. Since the project is a bit older I also had a look out for other projects and found Voxel51’s FiftyOne, which is like pix-plot on steroids. It allows the application of many more computer vision models and clustering algorithms, as well as a better finetuning and filtering of the results.
A takeaway of my exploration of computer vision models is the reflection on what features I actually need to extract from my dataset. I’m certainly not on the side of object classification and need something that is a tad closer to low-level feature extraction. This could also be a finding that I need to discuss with somebody a bit more knowledgeable of these subjects.
Tried to adjust the pixplot clustering hyperparameters but that needed another 4 hours.
Meanwhile I came up with a way of reducing the dataset size through removing similar images on a per-game basis, which removed roughly 42% of screenshots. I will try to work with this reduced dataset in pixplot and fiftyone.
pixplot with --min_dist 0 made practically the same clustering
2024-07-04
Moved some video game magazines for the Ludens archive from UNIL to the Gamelab in Bern.
The clustering is pretty good already and seems to take shape into a direction with which I can work with. I did some first explorations and the clusters look interesting. There is already a trajectory from text to image with mixed interfaces in between. There are aspects of colors and textures to be explored. Fiftyone is clearly the better choice to explore the corpus in that direction. I’m ready to run a clustering on a larger sample next week. 5’000 is less then 1% of the current corpus size (73’653 screenshots in total). So I would go for 10% next week. I might also see if the FiftyOne Similarity calculation is a better fit then the similar-image-remover script.