joshvm

17 ms·

211.
by joshvm 3y ago
Both Google and Apple have implemented models that aim to take the "best" picture from a video sequence, like a live photo. Have a look at MUSIQ and NIMA. https://github.com/google-research/google-research&#x2
212.
by joshvm 3y ago
Hue bulbs and strips run fine on HA. For a couple of reasons, I have a completely offline system running in a Docker container with a ZigBee dongle (mainly I have no control over my LAN). I don't use the bridge any more. You don't
213.
by joshvm 3y ago
The headline feels like clickbait. Does MLPerf also benchmark power consumption/cost per submission? It's not particularly illuminating that if you use a 10k GPU cluster, you can train things quickly and a "per chip" com
214.
by joshvm 3y ago
I have a couple of Tile fobs from before Apple moved into the space. They're still good enough for checked bag tracking while flying (at least to confirm the bag has made it to the terminal). Tracking is pretty limited without the subs
215.
by joshvm 3y ago
Automated/template-free astronomical spectrograph/spectrometer calibration. Given an emission spectrum and a list of known wavelengths that you expect to see, we try and determine the pixel/wavelength relationship. Traditiona
216.
by joshvm 3y ago
I would agree on one aspect though - deploying a model at scale is much closer to SWE than it is to foundational ML research. At a high level, you're deploying a function which has some known compute requirements. It requires setting u
217.
by joshvm 3y ago
GLP is a Monsanto-fronted publication [0], so there is a significant bias and I'd take anything they say (i.e. a company that really would like you to buy crop treatments) with a pinch of salt. [0] https://usrtk.org/ind
218.
by joshvm 3y ago
It's complicated, but John Oliver did a good segment on it. https://m.youtube.com/watch?v=6p8zAbFKpW0 Planting trees is not an easy solution. https://e360.yale.edu/features/phantom-forests-tree-pla
219.
by joshvm 3y ago
For what it's worth the best printer I've ever used (at least the most reliable) was a Markforged and that was very heavy lock in. Own brand filaments, cloud-only slicer, super expensive consumables. But that thing really did work
220.
by joshvm 3y ago
I think the suggestion is to try to collaborate with local universities. From experience, a lot of research/fieldwork in Africa is only practical with local experts and communities. Also look into attending or supporting the next Deep
221.
by joshvm 3y ago
I mean it actually happens. Not powering down the actual machine, but in my last house the landlord had a cleaner visit every fortnight. Occasionally I came home to find downloads broken or my remote connection would drop randomly in the af
222.
by joshvm 3y ago
The UK does fund SMEs, that's the Innovate UK program. It's a good pot for domestic money and relatively easy to get (or used to be, you'd see the same companies at every conference). It has money for solo industry and indust
223.
by joshvm 3y ago
Re satellite. At Sentinel resolution (10-20m) not much, maybe enough to distinguish plantation from natural forest spectrally. At Planet (3m) and below you can start to see large individual trees. It's very difficult to accurately meas
224.
by joshvm 3y ago
The literature is not particularly confident on the benefits of breathwork. A recent meta study found limited positive evidence. They conclude it can probably help some people, but the mechanism is unclear (and I read that as could be place
225.
by joshvm 3y ago
Yeah this is a good point. I've done some work with local experts to do tree ID from drone images over rainforest and there were several species where they would need to see the underside of the leaves to get better than the family.
226.
by joshvm 3y ago
Trees are generally difficult to classify well with computer vision. It's hard for the models to establish context because at a scale where you can see the whole tree you tend to include lots of background. If you include a bark photo,
227.
by joshvm 3y ago
Maybe a related question, but how good a proxy is cuff pressure for a gold standard method that measures inside an artery? Whenever I've had BP measured for clinical studies (not focused on BP specifically) they used a mix of cuff and
228.
by joshvm 3y ago
Australia and New Zealand have probably the best coffee (on average) of anywhere I've been, if you like milk drinks. The UK has a lot of fantastic coffee shops, but it's still overrun with Costas and a lot of places have added fla
229.
by joshvm 3y ago
Does Plenoxels also count? I believe they also use a similar approach https://alexyu.net/plenoxels/
230.
by joshvm 3y ago
I'd argue as a hobbyist you can get away with "block level" design - understand enough to know what you need, read datasheets and implement a reference design. For example making a stepper motor do what you want - probably no
231.
by joshvm 3y ago
Detection is via quite a simple method if I understand correctly - in the paper, section 7. It's basically a big camera. It's hundreds of stacked layers of tungsten and photo emulsion. Big heavy stuff because it (presumably) has a
232.
by joshvm 3y ago
You can buy known objects (expensive, for metrology calibration) or nowadays you could 3D print test objects. A simple one is to set up a flat board with white noise printed on it and then measure depth noise as a function of distance (eg f
233.
by joshvm 3y ago
> despite the fact that a depth map is the easiest way to hide poor quality I did my PhD on stereo and LIDAR; also in industry building these things. This is something that really annoyed us about camera companies selling stereo systems:
234.
by joshvm 3y ago
You can get an Orin with 8GB. Or the 32/64GB AGX module for 1.3/2.3k (3k for the dev kit). Not cheap but plenty of ram and 60W power target maximum. You could maybe run something on the 4GB Jetson Nano? But very slow: https:&#x2F
235.
by joshvm 3y ago
While these are great projects, they're for very different audiences. Code is a really well written pop-science book that goes into logic circuits. The Nand2Tetris book comments for example that: > The HDL that we will use is docume
236.
by joshvm 3y ago
I find OsmAnd to be OK in a pinch. I use it a lot abroad with full offline maps (the fee is less than $10 a year, bargain). You're right that search is bad. Often an exact search for a POI doesn't work, and it only shows the &quot
237.
by joshvm 3y ago
I'm not sure this is a correct assessment. The baseline ViT model ( https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M ) is presumably trained on a much larger dataset from HLS ( https://hls.gsfc.nasa
238.
by joshvm 3y ago
You can make it yourself - all the parts and boards are on GitHub (so also a group buy for assembled boards is an option). I couldn't find a BOM cost estimate though and I guess it'd be $1-200 from scratch, plus things like custom
239.
by joshvm 3y ago
To add a little to the existing replies. When you buy surface mount parts, they normally come on tapes - little plastic/cardboard strips with indents that the components sit in and a film over the top. These tapes are sold as reels if
240.
by joshvm 3y ago
It's more realistic to say that covid caused logistics issue two summers in a row. The first summer was presumably 2020 when almost every country locked down. In 2021 things weren't much better and the various polar programs were

More ›