stephanst

new dataset, new classes and model is yolov8 because most people prefer it. Also open weights.

Yep exactly. Commercial satellite inagery isn’t great for detecting anything the size of a tank (yet) but drones will get you there. The real difficulty for detecting anything like this is the training data, IMO that’s the interesting part of the problem. The actual AI models are democratized and getting better every day, but few people focus on the training data. I have my own pipelines for generating data that I have been improving over >3 years, which I have used for commercial clients to make AI systems for specific use cases. WALDO is a way of dogfooding those data synth methods. I am constantly toying with the idea of opening up the data synth stuff as a service company or as an API, but my feeling is there is not quite enough demand yet to warrant that. This is changing very quickly though and now might be the right time…

generalisation is 90% of the problem yep. This model is trained with a lot of augmentations and quite diverse data, but it still really needs fine tuning to specific use cases to get great efficiency unless you don’t care about false positives.

you don’t need AI for that, classic computer vision will work fine ;)

it is being used a lot for disaster response and traffic monitoring yep. Also surprisingly to me for analytics on utilisation of all kinds of infrastructure y trading firms. One of the bigger motivations to make it free was to see where people use the base model to understand whether there is a market there for a fine-tuned / more specialized version.

Some of the design choices of YOLOv7 make more sense to me in the choices of default augmentations and the structures of the very large versions of the networks. I find I can push it to marginally better recall. It’s slower than Ultralytics’ V8 but if you want to do stuff like offline processing of satellite imagery for instance or get 1fps on occupancy of a parking lot that kind of performance really doesn’t matter.

that class never really worked and has been removed from the new version of WALDO FYI, it’s not a military thing and shouldn’t be used as such

this was >50% of the motivation

Hey, thanks for posting. New release is coming tomorrow on HF BTW. AMA

looks cool, thanks !

I’d love to try it ! TBH this app was made in less than a day as a proof of concept: my real work focus is on AI for drones and satellite imagery. If you’d like to build an alternative with Qdrant I’d be happy to merge it into the repo!

Hey HN people, I'm releasing the source code of my little streamlit app called GPTflix that allows you to interact with >450K movie reviews and plots using vector search and the OpenAI LLMs. Enjoy!

url forwarding is broken, you can use https://www.gptflix.ai

And... it's back online, supported by Ben's Bites, my favorite AI newsletter! https://bensbites.beehiiv.com/

Sorry it appears that HN doesn't accept the format of the link somehow... it should be here https://gptflix.streamlit.app/

Hey HN. I made this last week for fun. It's using Pinecone as a vector database / backend, and context injection + data retrieval via embeddings search to power the chat function! It will only answer about movies that it knows (about 10K at the moment which pretty much maxes out the "small" tier Pinecone DB), but it will answer factually without making up fake information (which is the whole point of building a system like this instead of just using vanilla chatGPT). It has a hard time remembering context sometimes, so you might want to hard reload the site after a series of questions if you want to change subject. I'd love to head any and all feedback!

Hey HN. I made this last week for fun. It's using Pinecone as a vector database / backend, and context injection + data retrieval via embeddings search to power the chat function! It will only answer about movies that it knows (about 10K at the moment which pretty much maxes out the "small" tier Pinecone DB), but it will answer factually without making up fake information (which is the whole point of building a system like this instead of just using vanilla chatGPT). It has a hard time remembering context sometimes, so you might want to hard reload the site after a series of questions if you want to change subject. I'd love to head any and all feedback!

The dataset is a bit of a mess: there are labels as well as masks and they are not split up correctly... but still looks pretty usable with a good augmentation stack.

42 image batch size is scaled to the input layer of my DNN, otherwise yes :)

the 90% / 10% image was not a real world stat, it's supposed to be an illustration... but not well worded I'll grant you that.

What you are doing is the same type of work that I would put in the "pro user" bucket like ML. Actually I do the same thing with the same cards to render synthetic dataset for my ML pipeline, including using UE :)

Yep, this is a very valid point and I need to look more into this... which means rebuilding a lot of my toolchain but I think it would ultimately be worth the time investment!

Unfortunately yes... it looks like Nvidia have caught on to the fact that the consumer market doesn't need more VRAM ;)

Exactly.

> Thanks, this is exactly what I meant in a much better formulation :)

It seems like I've touched a nerve with a lot of people on here by saying "games don't need more powerful GPUs"... let me explain.

My contention is not that it's impossible to use more power in games, or that developers are not working on games that will be able to fill all that compute power... but that higher quality graphics are no longer pulling the gaming industry forward like what was going on 20 years, and also that improvements in graphics are more and more about AI enhancements to image quality and less and less about having the power to push more vertices around in real time.

The market for video games doesn't really care about the top 0.5% of PC gamers that can afford to buy the latest GPU from Nvidia every 2 years. The PS4 is still outselling the PS5, if graphics power was the main criterion for the actual market of video gaming that could not happen...

There's a lot of interesting discussion here!

I am the ex-CTO and co-founder of a company that does 3D-scan-based fit, specifically ML-based morphology extraction from 3D data of people (www.treedys.com , not a YC company ;-) ).

A lot of the points raised here in the comments are very valid: there is absolutely no consensus on standardisation of "fit"... but even if there was: "fit" is a very complicated concept that means different things to people in different settings. Furthermore there is often no consensus even inside big clothing companies as to which grading system (this is the "fit catalog" so to speak) is to be used for a specific collection / year / style etc... and those companies that do have this kind of thing standardised are often fighting with internal staff that grew up in a non-digital era to keep the documentation up to date, make sure the right files are shared, push people towards using digital design tools consistently etc...

But this does not mean - by any stretch - that the current system cannot be improved! There are many things that can be done with morphological data that are not immediately obvious in terms of improving fit for individual end users. Comparing "tshirt measurements" VS "human measurements" is the obvious starting point, but this is not actually the best way to make real-life recommendations, and there are many better systems that are possible once you have high quality metrics! Here is one for example: what about completely disregarding the "extracted body measurements in inches", and just building a flywheel where you take people's entire set of say 250 measurements as a vector, and use this data point along with a set of items they have bought (for themselves) and not returned. Once you have this kind of data for, say, 5000 people and 50 SKUs of items of clothing... you can build a very granular recommendation system, based on morphology, which will take into account many aspects of "fit" that are not immediately obvious in the measurements...

One thing is for sure: the big incumbents in fashion retail have felt the sting of shifting to an online-first world. In 2020 the total dollar amount of merchandise returned from online retail doubled (https://nrf.com/media-center/press-releases/428-billion-merc...). Literally: double the volume in one year. This was especially impactful in online fashion, where returns have always been a giant cost and a limiting factor to expansion ... but up until 2020 fashion retailers had been delaying action while desperately trying to modernise other aspects of their businesses, and morphology data felt like "too big of a problem for now.

Now, in 2021, with a few big players like Amazon taking the lead in "morphology-based fit" the situation is changing... and even the slowest fashion brands are starting to realise that not having a fit / morphology strategy in 2021 is starting to look like an automotive manufacturer NOT having an EV strategy in 2012 when Tesla was releasing the model S...

We're feeling it: after hanging on by the skin of our teeth for 6 years building deep tech solutions ( 3D ML to map a naked body on messy point cloud data of a clothed person for example, that was a tough nut to crack!), we're suddenly getting a lot more inbound from people we could not get on phone for a 5-minute conversation 2 years ago. And we finally have our first public-facing deployments in-store, using the full capture-to-ML stack (all developed in-house) that we have been crafting all this time! It feels good to finally get a product out and start getting the feedback loop going after so long.

All of this to say: congrats to Drapr, They made a great product and their timing was excellent! Let's hope this helps GAP achieve it's sustainability goals and reduce returns and waste, because in the end this will be the major benefit to all of us :) But in any case: morphology-based-fit is now "an idea whose time has come", and we can expect a lot of cool new things to come from this over the next few years!