About eight years ago I was trying to stream several videos of a drone over the internet for remote product demos. Since we were talking to customers while the demo happened, the latency needed to be less than a few seconds. I couldn't get that latency with the more standard streaming video options I tried, and at the time setting up something based on WebRTC seemed pretty daunting. I ended up doing something pretty much like JPEGs as well, via the jsmpeg library [1]. Worked great.
"worth the money" is hard to say, especially for devices like these where the value is not really so much on the plain features as much as a more subjective factors like the design and the UI. I would say that purely based on features - probably not, especially with post-covid pricing. There are more powerful iphone or android apps for much less. Behringer, and to some degree Korg and Roland offer lower-end devices for not much more that ultimately might be more useful and usable. But, I do own a couple of these little guys and they're fun. I wouldn't call them "jokes", but calling them "toys" - in the good and bad sense - would probably not be a stretch, even if you can get some nice sounds out of them. I used to keep a couple on my desk and just jam a little with them as a distraction.
There's a fun podcast by Arman Bohn/Distropolis (who himself has made some cool small-batch hardware synths) where he interviews makers of small hardware synths, https://open.spotify.com/show/30USGHPeGQ9ZyWQDyRnfcv might be of interest.
were you using single cycle waveforms or longer samples? in the former case, I guess there's not much to it, you just cycle through the waveform (in which case the waveform you choose would usually start and end at a zero crossing by construction, like sines, triangles, etc - or, if it does not, well, that will just create extra harmonics that might or not be desirable).
For the second case, it's more subtle. Most samplers that implement this feature will assume correct loop points (usually, but not necessarily at a zero crossing) are chosen manually by the user. Some of them implement cross-fading at the looping point to make that more forgiving, but that may be CPU/RAM intensive for some devices. If you're referring to small clicks you may get at the start and stop of sample playback, it's fairly common to use very short (ms or less) fade-in/fade-out to avoid that. There's a lot of books out there, but the main one I've read and enjoyed is this one, that happens to be free: https://cs.gmu.edu/~sean/book/synthesis/. it's more of a textbook than a cookbook.
Whenever I feel like doing that I just use "uv pip" and pretty much do the same things I'd do when using pip to messily install things in a typical virtual environment.
There's definitely a "rich get richer" effect for academic papers. A highly cited paper becomes a "landmark paper" that people are more likely to read, and hence cite - but also, at a certain point it can also become a default "safe" or "default" paper to cite in a literature review for a certain topic or technique, so out of expediency people may cite it just to cover that base, even if there's a more relevant related paper out there. This applies especially in cases where researchers might not know an area very well, so it's easy to assume a highly cited paper is a relevant one. At least for conferences, there's a deadline and researchers might just copy paste what they have in their bibtex file, and unfortunately the literature review is often an afterthought, at least from my experience in CV/ML.
Another related "rich get richer" effect is also that a famous author or institution is a noisy but easy "quality" signal. If a researcher doesn't know much about a certain area and is not well equipped to judge a paper on its own merits, then they might heuristically assume the paper is relevant or interesting due to the notoriety of the author/institution. You can see this easily at conferences - posters from well known authors or institutions will pretty much automatically attract a lot more visitors, even if they have no idea what they're looking at.
Ah, so that's why the camera seemed familiar, I have a couple of the luxonis cameras around the office :). Re: kiwi, those are good points. Thank you for the answer!
Hi, I am currently considering a Lekiwi build but I am intrigued by Mars. Outside of the need for external compute, what issues did you find with SO101 and Kiwi?
Also I am curious about a couple of the parts, if you don't mind sharing - are those wheels the direct drive wheels from waveshare? And what is the RGBD camera? (Fwiw, even if it's hefty the MARS price tag seems fair to me).
Seems like a corporate version of the "buy vs build" question. If it's true that the goal is to become more approachable to students and hobbyists (which personally I think would be a good idea) - then Qualcomm must've evaluated both options and decided "buy".
I feel like the raspberry pi pico is more of a competitor to the arduino than the raspberry pi - there's quite a few applications where having a whole linux operating system is a hindrance compared to running on bare metal, especially anything that needs real time control of signals. (Although you can get around this on the pi by connecting peripherals via USB/serial/i2c which themselves might use MCUs).
Then again, one of the more accessible (IMO) ways of using pi picos is with the arduino environment, or its cousin platformio. I do think that even if in some ways the arduino abstractions can be limiting in some ways, in practice it's often a big timesaver for more casual (and not so casual) applications. It gives you easy access to a large ecosystem of libraries across a lot of hardware platforms.
This is a great book - learned a lot from the first edition back in the day, and got the second edition as soon as it came out. It's always fun to just leaf through a random chapter.
Yeah, that's a good point. I was thinking in the US context, but I also have some experience for the academic evaluation process in Chile, and there were similar issues to what you were describing. The "bean counting" part of it was an issue for academics in CS, because the rules were the same across departments, even where it didn't really make sense. So for example CS profs got no credit (towards promotions) for publishing in conferences, even if they were highly selective ones like ICCV or N(eur)IPS.
Depends on who is doing the "careers valuing" and how closely they're looking. At a coarse level, especially for jobs in industry, venue is a pretty simple (but obviously imperfect) indicator for quality. If you've managed to publish one or more papers at the most selective venues (esp. as main author), then I would assume there's a decent chance you are good at research, even if I don't know anything about the subfield you work on. As a further indicator, the number of citations is also a noisy but easy to check proxy for "impact".
But for academic or other high-level research jobs, whoever is doing the valuing is going to look at a lot more than just the venue.
Just one data point, but if it's as nice to use as their open source tools and not outrageously expensive, I'd be a customer. Current offerings for private python package registries are kind of meh. Always wondered why github doesn't offer this.
I had a similar thought, but at the same time, if people were mandated to use Windows or MacOS then that would also pretty much lock you into their respective window managers. I guess it feels more restrictive partly because it's more common to pick and choose WMs on linux. (And partly because, yeah, seems like the setup goes way beyond just a distro+WM).
The few examples they show do look pretty good for a wifi-based method, although who knows how cherry-picked they are. I wonder how much the "SLAM" part is contributing and how sensitive that is to the sensor quality on the phone. I would've assumed that they'd be using vision, which seems to be the method of choice for other companies like niantic. The ground-truth data part for vision would certainly be more onerous, though.
This sounds pretty similar to CDE, which I see they cite in the paper. Back in the pre-docker days I remember using CDE a few times to package some C++ code to run on some servers that didn't have the libraries I needed. Pretty cool tool.
I was going to mention this sounds like the idea behind adversarial approaches, which I guess go all the way back to game theory and algorithms like minimax. They're definitely used in the control literature ("adversarial disturbances"). And of course GANs.
You can examine the code of zxing-cpp (which is fairly nice IMO) for a simple, "classical computer vision" approach to this. It's not the most robust implementation but it is pretty functional.
But in general, you can divide the problem more or less like this (not necessarily in this order)
1. find the rough spatial region of the barcode. Crop that out and only focus on this
2. Correct ("rectify") for any rotation or perspective skew of the barcode, turn it into a frontoparallel version of the barcode
3. Binarize the image from RGB or grayscale into pure black and white
4. Normalize the size so that each pixel is the smallest spatial unit of the barcode.
You can roughly divide barcode reading into a "frontend" and a "backend". The backend is the most well understood (but not necessarily trivial) part: you take a binary image, with each pixel corresponding to one little square in the QR code, and decode its payload. It doesn't need computer vision. The "frontend" is the part that takes the raw image containing the barcode and tries to find the barcode, and convert the barcode it finds into a nice, clean binary image for the backend. This is a computer vision problem and you can arbitrarily fancy, including up to using the latest trends in ML vision models. However, this isn't necessarily needed in most cases; after all, barcodes are designed to be easy to read for machines. With a large, sufficiently well focused and well exposed image of a barcode you can get away with simple classical computer vision algorithms like histogram-based binarization and some heuristics to identify the spatial extent of the barcode (for example, most barcode symbologies mandate "quiet space" (blank space) to be around the barcode, and have start and stop markers; QR codes have those prominent concentric squares on the corners).
As for implementation, Zxing-cpp [1] is still maintained, and pretty good as far as open source options go. At this point I'm not sure how related it is to the original zxing, as it has gone substantial development. It has python bindings which may be easier to use.
On mobile, Google MLkit and Apple vision also have barcode reading APIs, not open source but otherwise "free" as in beer.
I've found myself adopting this philosophy for a specific use case: monitoring ML training jobs. It's pretty common to see people output training metrics (loss, validation accuracy, etc) every N batches, iterations or epochs. And that does make sense for a lot of reasons, and it's pretty simple to do. But also when you're exploring models that might have wildly varying inference latencies, or using different hardware, or varying batch sizes, or using a differently sized dataset, all of those might end up reporting too infrequently to get an idea of what's happening or too frequently and just spamming too much output.
Checkpointing the model every N iterations/epochs/batches has a similar problem - you may end up saving very few checkpoints and risk losing work or waste a lot of time/space with lots of checkpoints.
So I've often found myself implementing some kind of monitoring and checkpointing callbacks based on time, e.g., reporting every half an hour, checkpointing every two hours, etc.
I don't know if there's an earlier source, but I'm guessing Matlab originally popularized the `imread` name, and that OpenCV (along with its python wrapper) took it from there, same for scipy. Scikit-image then followed along, presumably.
I'm not sure I quite understand what you're aiming at with these questions, but there are certainly techniques in ML based on thinking of functions as vectors. The first one that comes to mind is Anyboost [1], which views boosting as doing "gradient descent" in function space - where each "gradient step" is not a typical vector (as you'd see e.g in neural nets) but a function, corresponding in practice to a base classifier. Another that is very popular are gaussian processes - one way to think about them is as modeling functions as samples from an infinite-dimensional gaussian (at least some of them).
"why didn't we think of this sooner?", asks the article. Not sure who the "we" is supposed to be, but the robotics community has definitely thought of this before. https://robo-affordances.github.io/ from 2023 is one pretty relevant example that comes to mind, but I have recollections of similar ideas going back to at least 2016 or so (many of which are cited in the V-JEPA2 paper). If you think data-driven approaches are a good idea for manipulation, then the idea of trying to use Youtube as a source of data (an extremely popular data source in computer vision for the past decade) isn't exactly a huge leap. Of course, the "how" is the hard part, for all sorts of reasons. And the "how" is what makes this paper (and prior research in the area) interesting.
Yeah, that was my first thought. And it's not just about them accepting typst, but also whether they would provide a template using typst, like they currently do for latex. Using the conference/journal template to write the article saves a lot of time for both submitters and editors (who have to deal with hundreds, if not thousands of submissions).
Like the other commenter said, their retail price is $2000. In the used market, they are going for about $1400-1500 based on reverb.com stats. So at least compared to those prices, $1400 new is a decent deal. And for what it's worth, they sold out pretty quickly today (or so their online store says).
Of course, $1400 is still a hefty price tag, I won't argue about that (and there's not much anything new to say about that topic, considering it's like 80% of online discourse about TE).
Yeah, the article was painting with a bit too of a broad stroke IMO, though they did briefly acknowledge "special exceptions" such as satellite or medical imagery. It's very application-dependent.
That said, in my experience beginners do often overestimate how much image resolution is needed for a given task for some reason. I often find myself asking to retry their experiments with a lower resolution. There's a surprising amount of information in 128x128 or even smaller images.
There's been a huge amount of work on image transformers since the original VIT. A lot of it has explored different schemes to slice up the image in tokens, and I've definitely seen some of it using a multiresolution pyramid. Not sure about the RL part - after all, the higher/low-res levels of the pyramid would add less tokens than the base/high-res level, so it doesn't seem that necessary. But given the sheer volume of work out there I can bet someone has explored this idea or something pretty close to it already.
Slicing up images to analyze them is definitely something people do - in many cases, such as satellite imagery, there is not much alternative. But it should be done mindfully, especially if there are differences between the training and testing steps. Depending on the architecture and the application, it's not the same as processing the whole image at once. Some differences are more or less obvious (for example, you might have border artifacts), but others are more subtle. For example, contrary to the expected positional equivariance of convolutional nets, they can implicitly encode positional information based on where they see border padding during training. For some types of normalization such as instance normalization, the statistics of the normalization may vary significantly when applied across patches or whole images.