Full article: https://iliashaddad.com/blog/i-indexed-669-gb-of-my-gopro-videos-using-my-m1-max-computer
You might want to add something like yolo finetune to detect scenes + face recognition too.
I'm really bullish on taking more video of my kids, with the thought that it will become easier and easier for AI to put them into little compilations I can enjoy later.
I booted up my old PS3 from my uni days (20 years ago?) and found all of the music I had on it because I used it for everything at the time. Some seriously nostalgic music I'd completely forgotten about.
Google loves scanning stuff on in the cloud though.
Years from now they'll be getting "hey look at BIKE BRANDS' NEWEST CHEAP BIKE REMEMBER WHEN YOU USED TO RIDE BIKE BRAND BIKES"
M1 Max CPU is an ARM/SoC, comparable to an 11th gen Intel i9
ref: https://www.cpubenchmark.net/compare/4585vs4245/Apple-M1-Max...
- "unified" ram makes all the system ram available as VRAM - dedicated ai coaccelerator thingy
Both of these reasons allow the apple silicon chips to crush conventional cpus in these kind of AI model workload stuffs
No idea about what the windows arm stuff is capable of. I know they use Qualcomm snapdragon chips though.
comes with some nifty features like NLE- integrations, people search, MCP, API etc
Disclaimer: one of the co-founders
Other comments mention davinci resolve has this built in. How would you compare the two?
Her client was recording while committing the abhorrent crime. The criminal would otherwise have got off.
From my perspective, the GoPro camera produced a good outcome. Still, one has wonder why anyone to record their criminal actions.
She would rather have done corporate law but did not have the academic credentials or the networks needed for a job at the likes of Latham Watkins or White and Case.
Still it is good for society that criminals get the worst lawyers to defend them.
Aha, it makes total sense. This number sounds much more reasonable than “669 GB”, since the actual total size of processed frames would be like 10-30 GB.
(Not downplaying anything. Doing-at-home always requires some math on practicality)
> Total compute time 67h 40m 42s
I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
The reason why is “669 GB” is the total raw footage size when I'm doing the video processing, I downscaled each frame to 720p to make the video processing much faster and I don't need full original quality in order to get accurate results (as far as I know and experiment with).
> I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
For now, I found that NVIDIA GPU for example RTX 3060 with 12GB Vram was much faster than my M1 Max. (still working on optimizing for speed and accuracy).
But it's not as fun as running local model right here on your computer on your own desk. It feels like magic.
https://news.ycombinator.com/item?id=48222733 https://blog.simbastack.com/indexed-a-year-of-video-locally/
I wasn't familiar with your project though, interesting stuff.
I'm trying to add more photography related features to Framedex but yeah there's so much we can do locally, exciting times.
Good job for the article and the project. That's great, yes local models are getting better and better
I think Adobe premiere pro have it as well but cloud processed
Take a fast, small and powerful LLM running locally to index my personal data like images, videos, documents and enrich them and tag with the enriched metadata.
Want to group by people - Search tagged metadata and group it What to search an image by description - tagged metadata What to organize by anything - tagged metadata
This should (hopefully) put an end to my file clutter
Local LLMs sound so cool but I know they won't be easy to setup or use for common joe like me.
And once set up it's easy to use even for non technical people.
For the dog barking videos, those are only the video scenes that I have a dog barking sound in the video.
I'll keep adding more prompts and example videos, keep an eye for that
Did you ever visit crazyguyonabike.com? A long time ago I had the pleasure of following the journey of a friend of a friend of a friend on that site:
https://www.crazyguyonabike.com/doc/?doc_id=2405
Stuff like that I guess?
Yep. I had the same problem.
> Then, run the frame analysis pipeline [...] I have a face recognition plugin using my custom faces data, object detection, on-screen text, shot type, and scene description [...] we will have three vector DB collections that have all the information about our videos, like video location metadata, camera name, faces recognized, objects detected, on-screen text, transcription, description of each scene, and many more [...] we can get better indexed data if you use the advanced mode indexing to use the Qwen2.5-VL-7B-Instruct model to understand and describe your video much better, but at a slower indexing speed
Yeah, uhm... ok :)
If anyone else has a similar problem, the real solution is as follows:
1. When recording, if you witness an interesting moment worth saving later, press the power button — this will mark the current moment in the video as a chapter.
2. Find the chapters later when editing and cut them into clips.
3. You're done :)
This has two main benefits over the insanity above:
1. It's trivially simple instead of insanely complex and inefficient.
2. It will reliably catch all the stuff you find interesting, since you're the one doing the marking.
The downsides:
1. Doesn't work retroactively.
2. It may miss interesting stuff if you miss it at the time as well.
3. Only works for this use case.
4. Nerds won't salivate over your usage of cutting edge tech.
Being able to semantic search over your library is useful, but does it solve the review problem? I feel like you would still need to watch the footage back before you know what you're working with.
Frame level embedding it covering a lot, but can miss out on a lot of action related searches.
EDITED: I didn't realize Whisper was a local model. I never tried transcription before, so I had always figured it was a pay model by OpenAI. I'll have to check it out (although the runtime listed here is a bit daunting).
For that project I'll say I don't see much degradation in embedding quality at much much worse quality than 720p (all the way down to 240p), which speeds things up considerably. Although I don't really do face or object detection, just scene embeddings. To me any process whereby it would take longer to process the video than watch it is probably a no go in general. Obviously a challenge for local-first analysis.
The world and our discourse around it has changed so much over the past ten years and now with this kind of technology I'm so excited to be able to classify these images from my iCloud and start on the project.
I might be better off getting something with a beefy GPU on AWS or Google cloud.