Ok, this cracks me up. I have a little company where we built models to identify vegetation species data from drone and satellite imagery across a target site. It's very funny, because segmenting...
Ok, this cracks me up. I have a little company where we built models to identify vegetation species data from drone and satellite imagery across a target site. It's very funny, because segmenting types is challenging, but without a huge amount of training data identification is just impossible. I went into it really naively thinking this would be an easy issue in 2018.
As ongoing R&D we always test new models coming out to see if there are any improvements from pre-trained multimodal models like Gemini and more specifically for image segmentation/identification (because Gemini is garbage at anything that doesn't pop up in a capcha prompt) Nano Banana. And it's still awful. I'd love to know what backend this runs on with those predictive numbers. All of the sample content they showed was doing a pretty poor job of IDing plants. And that's just plant v tree, which in principal should be suuuuuper easy.
I was curious if this is someone's dissertation based on the level quality and boom - Lukas Pilka's PhD from the Academy of Arts Architecture and Design in Prague. No shade to PhDs, nearly all my local friends are in or have recently finished theirs, but always a weird air of "official" or "professional" but often end up feeling empty? In my past career I worked with a bunch of Universities on interpretive digital works with art and heritage and all the projects came out feeling like this. Largely looking impressive but having very little intractability or just a weird quality issues.
Anyway, sorry for the long rant. It just spoked to a weird hybrid of my career thus far!
This one I'm sharing because it's so bad it's actually kind of thought-provoking. The AI curation that I got was Plants motif + Trees motif. Very nice collection of art, Mr. Computer. Makes me...
The Digital Curator application allows you to explore the art collections of Central European museums and search for artworks based on specific motifs.
Users of the application can build their combination of objects and reveal how often the subject has occurred across the centuries, view graphics, drawings, or paintings that represent it in different epochs, and compare data with other themes.
The Digital Curator offers a quantitative view of cultural history based on the frequency of symbols and iconographic themes in many artifacts, not on a detailed observation of individual items. This distant viewing can be especially useful if our interest is aimed at exploring a genre, rather than a specific work, to understand the overall social conditions, rather than the life of a particular artist, or to interpret the overall political situation, rather than the views of the selected author. Exploring big cultural-historical data may bring new insights into abstract social phenomena such as cultural and economic influence, canon issues, the relationship between the center and the periphery, or the functioning of the art market. It can also help us better observe the migration of motifs and their takeover across centuries and distant regions.
The Digital Curator database now contains 196,116 works from the collections of 91 museums from Austria, Bavaria, the Czech Republic, and Slovakia.
This one I'm sharing because it's so bad it's actually kind of thought-provoking. The AI curation that I got was Plants motif + Trees motif. Very nice collection of art, Mr. Computer. Makes me feel very Daft Punk - Humans After All. The extra sigfigs that occasionally slip through really sell it.
I'm 56.00000000000001% sure that's a plant.
I'm 56.99999999999999% sure that these pieces fit together conceptually and thematically--which is to say, I'm not sure.
edit: Portrait of the Musician is perhaps not so bad. Maybe I just have a dabbler level of musician portrait comprehension though... I think it also helps to reign in the date range to a specific era--I'm surprised the randomization doesn't attempt this but you can manually set it via URL param
Ok, this cracks me up. I have a little company where we built models to identify vegetation species data from drone and satellite imagery across a target site. It's very funny, because segmenting types is challenging, but without a huge amount of training data identification is just impossible. I went into it really naively thinking this would be an easy issue in 2018.
As ongoing R&D we always test new models coming out to see if there are any improvements from pre-trained multimodal models like Gemini and more specifically for image segmentation/identification (because Gemini is garbage at anything that doesn't pop up in a capcha prompt) Nano Banana. And it's still awful. I'd love to know what backend this runs on with those predictive numbers. All of the sample content they showed was doing a pretty poor job of IDing plants. And that's just plant v tree, which in principal should be suuuuuper easy.
I was curious if this is someone's dissertation based on the level quality and boom - Lukas Pilka's PhD from the Academy of Arts Architecture and Design in Prague. No shade to PhDs, nearly all my local friends are in or have recently finished theirs, but always a weird air of "official" or "professional" but often end up feeling empty? In my past career I worked with a bunch of Universities on interpretive digital works with art and heritage and all the projects came out feeling like this. Largely looking impressive but having very little intractability or just a weird quality issues.
Anyway, sorry for the long rant. It just spoked to a weird hybrid of my career thus far!
This one I'm sharing because it's so bad it's actually kind of thought-provoking. The AI curation that I got was Plants motif + Trees motif. Very nice collection of art, Mr. Computer. Makes me feel very Daft Punk - Humans After All. The extra sigfigs that occasionally slip through really sell it.
I'm 56.99999999999999% sure that these pieces fit together conceptually and thematically--which is to say, I'm not sure.
edit: Portrait of the Musician is perhaps not so bad. Maybe I just have a dabbler level of musician portrait comprehension though... I think it also helps to reign in the date range to a specific era--I'm surprised the randomization doesn't attempt this but you can manually set it via URL param