Oh thank you! Hm this still looks super glossy (and/but deeper) when fused to me. Similar in VR I get the illusion strongly for a 2d or 3d object, but goes away if I can move the object (participants don't agree with me on that last bit)
We haven't asked about relief, only gloss. Wd be interesting.
We haven't asked about relief, only gloss. Wd be interesting.
June 20, 2025 at 9:56 PM
Oh thank you! Hm this still looks super glossy (and/but deeper) when fused to me. Similar in VR I get the illusion strongly for a 2d or 3d object, but goes away if I can move the object (participants don't agree with me on that last bit)
We haven't asked about relief, only gloss. Wd be interesting.
We haven't asked about relief, only gloss. Wd be interesting.
Oh hey @maartenw.bsky.social good to see you on Bluesky! Check out the poster above. Your gloss illusion works well in VR and is strongly geometry dependent. But our Ss persist in seeing it even when they have stereo info and can move the object. Planning expt v3 where I hope to finally break it!
June 20, 2025 at 6:58 AM
Oh hey @maartenw.bsky.social good to see you on Bluesky! Check out the poster above. Your gloss illusion works well in VR and is strongly geometry dependent. But our Ss persist in seeing it even when they have stereo info and can move the object. Planning expt v3 where I hope to finally break it!
It was actually a conversation with Sarah Weigelt a year ago about that patient which gave us the idea for this project!
Sarah was asking what kind of DNN model might show similar deficits and I was like "....most of them, I reckon?"
Sarah was asking what kind of DNN model might show similar deficits and I was like "....most of them, I reckon?"
April 17, 2025 at 9:54 PM
It was actually a conversation with Sarah Weigelt a year ago about that patient which gave us the idea for this project!
Sarah was asking what kind of DNN model might show similar deficits and I was like "....most of them, I reckon?"
Sarah was asking what kind of DNN model might show similar deficits and I was like "....most of them, I reckon?"
One way of looking at these results is as a snapshot of which visual tasks are easy vs hard to learn from massive human-generated web corpora of text-and-image pairs.
It's interesting too that human vision, trained on natural images, generalises effortlessly to simplified lines, gabors, and shapes.
It's interesting too that human vision, trained on natural images, generalises effortlessly to simplified lines, gabors, and shapes.
April 17, 2025 at 8:35 AM
One way of looking at these results is as a snapshot of which visual tasks are easy vs hard to learn from massive human-generated web corpora of text-and-image pairs.
It's interesting too that human vision, trained on natural images, generalises effortlessly to simplified lines, gabors, and shapes.
It's interesting too that human vision, trained on natural images, generalises effortlessly to simplified lines, gabors, and shapes.
tldr: VLMs match or beat human performance in what vision science considers complex, high level visual tasks (face and object recognition) but struggle with low and mid level tasks (length, orientation, texture, occlusion).
We see "clinical deficits" (<2SD below human norm) on most "simple" tasks.
We see "clinical deficits" (<2SD below human norm) on most "simple" tasks.
April 17, 2025 at 8:26 AM
tldr: VLMs match or beat human performance in what vision science considers complex, high level visual tasks (face and object recognition) but struggle with low and mid level tasks (length, orientation, texture, occlusion).
We see "clinical deficits" (<2SD below human norm) on most "simple" tasks.
We see "clinical deficits" (<2SD below human norm) on most "simple" tasks.
All plants, left long enough, will tend towards a brassica.
This is the way the world ends — not with a bang, or a whimper, but with a crab atop its cabbage throne.
This is the way the world ends — not with a bang, or a whimper, but with a crab atop its cabbage throne.
January 20, 2025 at 1:02 AM
All plants, left long enough, will tend towards a brassica.
This is the way the world ends — not with a bang, or a whimper, but with a crab atop its cabbage throne.
This is the way the world ends — not with a bang, or a whimper, but with a crab atop its cabbage throne.
Phrased very broadly there. But they're probably not wrong that providing input to a CNN in the form of nicely structured receptive field responses is better than random RF responses (their random control) or raw pixels (their alex/google nets), at least when you have v limited training time & data.
December 16, 2024 at 3:54 AM
Phrased very broadly there. But they're probably not wrong that providing input to a CNN in the form of nicely structured receptive field responses is better than random RF responses (their random control) or raw pixels (their alex/google nets), at least when you have v limited training time & data.
Devil's Finger - a stinky native fungus!
December 16, 2024 at 2:13 AM
Devil's Finger - a stinky native fungus!