Do you have any resources you would suggest outside the links CVAT?

[parrotmaker]

Do you have any suggestions for resources other than CVAT's links? I looked into installing CVAT locally, and it seemed to be a good choice. What you suggest about feeding the photographs into a folder and then using detect.py and Mike's model as default has me a little confused owing to my lack of knowledge, but I believe you are getting at what I want to learn how to achieve. Would this be something I could do with the built-in CPAI training module, or do you mean running the training code directly in a VSCode-like environment?

[TimG]

Hi, is this a response to a thread on a different forum ?

[parrotmaker]

Hi, is this a response to a thread on a different forum ?

No. I'm researching the issue so I wanted to ask.

[BBK]

So you're trying to train your own models? Fair warning, it's a bit of a PITA. Mike's models work for 90% of use cases and will probably work better than anything you train yourself unless you get a BUNCH of images. He trained his on the entire COCO dataset I believe. There's a ton of stuff randomly "documented" on the forums-which-should-not-be-named, which I won't link to as that community is fairly toxic and I won't give them the traffic. Anyone can also grab the COCO dataset, Open Images Dataset, or any number of others if you want to start with that. These datasets (and associated training) aren't the easiest things to work with though if you don't already know what you are doing or are willing to put in quite a bit of research. I've put days into training my own image set (over the course of months, as a side project) and I still don't understand half of the workflows that seem to be "standard".

I did pay for CVAT, rather than use the free/locally hosted version, both because I hate docker that much and to remove the limitations. Basically CVAT is just an annotation tool. The general idea is that you upload a bunch of images, manually label some of them, train a model and use that to automatically label a bunch more (workload reduction), clean those up, train a model, auto label more, train a model, etc. until you get enough tagged images that a model works reliably. I'm mainly stuck on the automatic labeling myself. CVAT currently only supports custom models from Hugging Face or Roboflow. I don't like Roboflow's public only requirements, and I haven't figured out Hugging Face yet. I can train models locally, use that to annotate outside CVAT, and upload those. It's a pain to keep everything organized that way though (working outside CVAT that is).

The TLDR is that, right now, it's not easy to create custom models. If you want to, you'd better be ready to start Googling.

[parrotmaker]

So you're trying to train your own models? Fair warning, it's a bit of a PITA. Mike's models work for 90% of use cases and will probably work better than anything you train yourself unless you get a BUNCH of images. He trained his on the entire COCO dataset I believe. There's a ton of stuff randomly "documented" on the forums-which-should-not-be-named, which I won't link to as that community is fairly toxic and I won't give them the traffic. Anyone can also grab the COCO dataset, Open Images Dataset, or any number of others if you want to start with that. These datasets (and associated training) aren't the easiest things to work with though if you don't already know what you are doing or are willing to put in quite a bit of research. I've put days into training my own image set (over the course of months, as a side project) and I still don't understand half of the workflows that seem to be "standard".

I did pay for CVAT, rather than use the free/locally hosted version, both because I hate docker that much and to remove the limitations. Basically CVAT is just an annotation tool. The general idea is that you upload a bunch of images, manually label some of them, train a model and use that to automatically label a bunch more (workload reduction), clean those up, train a model, auto label more, train a model, etc. until you get enough tagged images that a model works reliably. I'm mainly stuck on the automatic labeling myself. CVAT currently only supports custom models from Hugging Face or Roboflow. I don't like Roboflow's public only requirements, and I haven't figured out Hugging Face yet. I can train models locally, use that to annotate outside CVAT, and upload those. It's a pain to keep everything organized that way though (working outside CVAT that is).

The TLDR is that, right now, it's not easy to create custom models. If you want to, you'd better be ready to start Googling.

Thanks! I just want to learn more about CVAT.