DailyPapers: Hippocampal fields and hand tracking

 The brain research paper I read today was "Imaging Human Medial Temporal Lobe with High-Resolution fMRI", Carr et al, Neuron, 2010, https://www.cell.com/neuron/fulltext/S0896-6273(09)01037-X

Figure from: https://www.cell.com/neuron/fulltext/S0896-6273(09)01037-X

 I wanted to talk about a paper that this paper cited. It was Bakker et al, 2008. They used a repetition suppression task to study pattern completion and separation in different sub-modules of the hippocampus. Repetition suppression refers to a neural phenomenon where stimuli that are familiar evoke smaller neural responses than novel stimuli. I actually did a little project after a Princeton neuroscience summer school on this paradigm, trying to model this phenomenon on a cellular level https://github.com/mariakesa/NeuralSimulations/blob/master/NAND2016Report_MariaKesa.pdf .  

 Anyway, Bakker et al showed participants repeated items, novel items and "lure" items that resembled previously shown objects but were not exactly the same. They found that neural activations were greater in response to novel and lure trials in DG and CA2/3 areas of the hippocampus. This means that novel and lure items were treated as the same for these circuits, implicating these areas to pattern separation (i.e. the circuits were able to discriminate that they have not previously seen the lure items). On the other hand, CA1 and subiculum areas of the hippocampus were greater only for novel stimuli, meaning that repeat stimuli and lure stimuli were treated as the same. This means that these brain areas are biased toward pattern completion. 

 

I think repetition suppression is an important neural effect in advertising. Novelty is an important driver of culture and cultural fascination. The above Abidas logo is a lure, a rip-off of the original. The human brain will process this stimulus differently than the original. 

The computer science paper I read today was from Facebook research, "Constraining Dense Hand Tracking with Elasticity", Smith et al, SIGGRAPH, 2020, https://research.fb.com/publications/constraining-dense-hand-surface-tracking-with-elasticity/(there's also a video).

 From the paper: "In this paper, we present a new method to track dense hand surfaces to a high degree of fidelity from multi-view image sequences using a physically based model. Specifically, we constrain the solution space of a vision-based tracking algorithm with an elastic volume deformation model and a collision response model, regularizing the entire hand geometry and deforming occluded regions of the hand stably and plausibly. The remaining visible regions are tracked based on visual data from multi-view cameras. To the best of our knowledge, our method is the first to track details such as creasing, bulging, and deformations under extreme self-contact and self-occlusion for one and two hand motion sequences."

So cool!

They use optimization to navigate the solution space. Specifically, they minimize an energy functional consisting of three terms: a vision term, a physics term and a link term that matches the mesh model with the visual data. The vision term evaluates the degree to which the camera data is explained by the 3D surface estimates i.e. the mesh model. I'm guessing it's a bit like regression and variance explained. The physics term is a Neo-Hookean elasticity model that penalizes volume changes. They also have a collision detector that uses ray tracing. It's cool they also mention how they accelerate the ray-surface intersection tests. They use a fast spatial hash and a kd-tree.

 That's it folks!