(source)
You own a 3D Printer? You're using OctoPrint? Your printer is equipped with a camera? You want to help in develop a deep nerual network to detect failing 3D prints without looking at your printer all the time? You're willing to donate frames of your camera to a public dataset?
Then you're the right person. Please install the plugin and help us to make a better world! (just kidding, I'm just tired of failing prints in an overnight print or when I'm at work).
You can download the plugin here (I'm sorry - you can't as of now 😥)
Are you tired of failed prints during the night? Tired of clogging, spaghetti, warping and objects not sticking to the bed?
I started to look for object vision based solutions to monitor my prints or open datasets to train my own model without any success. There are several projects available like The Spaghetti Detective, Kronos Data Collector, etc. but I was searching for a more open and extensible approach to solve this problem. But unfortunately everything starts with a good dataset 😣. I labeled a few hundred pictures of my beloved Prusa MK3S and trained several models to detect certain edge cases which worked out quite well but I want to find a generalized solution which can be used accross the 3D-Printing community. So let me just start with some core tenets of this project:
- Everything will be open-source!
- The dataset is crowdsourced and licensed as ODbL
- Labeling of the data will be crowssourced, too.
- I'm very open for contributors, feel free to reach out to me, create a PR, create an issue, whatever necessary to make this project successful!
Okay, but what is this project actually doing?
Currently there is nothing more then a data collector, implemented as OctoPrint plugin, which takes a picture every x-seconds and sends it to a remote API. The code itself is pretty messy and subject to be improved (feel free to leave a PR!).
TLDR: Where can I download the plugin?
WIP: The plugin is not yet published, I'm working on it...
Where will my data be stored?
The collected data will be stored in Frankfurt (DigitalOcean)
What happens if I'm accidentially naked in front of my 3D printer?
The collected images will be cleaned in a unsupervised fashion. I'm using a simple model to detect human-beings and NSFW content within the pictures (and delete them...)
But why can't I see any machine learning models in this repository? 😱
For now only the data collector plugin is available because as you know - every good model starts with a good dataset 😝
Hmm, when I install the data collector, what will happen with the pictures?
By installing and enabling the plugin, you agree to publish the images to the public licensed as ODbL. The data collection is anonymously and the data will be kept indefinitely.
Where can I find the pictures in this repository which I published yesterday?
The dataset will be cleaned and distributed at least once a month.
I want my data to be deleted because I printed xxx which I don't like to be part of a public dataset?
The plugin writes a unique hex to your octoprint.yaml configuration file. Feel free to create an issue as fast as possible (before I publish the new version of the dataset) to get your data cleaned.
Why is the plugin licensed as GNU AFFERO GENERAL PUBLIC LICENSE Version 3?
Because I borrowed some code from the Spaghetti Detective plugin (nice work btw!)
Why has this README so many spelling and grammar issues?
Sorry, it is late and I'm not a native speaker - feel free to create a PR (please)...
How can I be added in the list of contributors?
You're contributing to the dataset by using the collector plugin or contributed to the code? Just create a PR and add your name. Wait what - but how are you checking that people are really using the plugin? I don't and I don't intend to do so!
Will this be a cloud service or what are your plans?
I intend to create a lightweight model which can be used on Jetson or similar. However at the beginning I will probably experiment with deeper models to find a good baseline.
How do you want to make this possible, do you actually have any idea of deep learning?
Yes I have but this is not subject of this repository. I already explored several approaches using pixel-wise image segmentation to detect the print, the bed and the print head. This helps to detect separations from the print bed, extereme clogging and gives us a perfectly trained feature extractor backbone which can be used to explore certain other problems like spaghetti (Which can be easily part of the segmentation itself), warping, etc. I ended up exploring a multi-task learning approach with different losses accross the tasks and different weights for each loss. However this approach was highly complicated but I think it is feasible to use a big network to train a lightweight network in a semi-supervised approach to run it at the edge.
What framework do you prefer?
Definitely PyTorch! 😍
Can I contribute to the project?
Yes definitely (and I would highly appreciate it)! There are so many topics to work on, just to name a few:
- Improve the code - it has no tests, it's pretty hacky, not documented and maybe not using the right architecture.
- Correct my spelling mistakes. 🙃
- Automate some things (GitHub Actions, Dependabot, ...)
- Make all your friends to install the collector plugin
- ... and many more ...
What exactly will be tracked?
- A unique identifier (hex format) which gets generated when you install the plugin the first time and gets stored in the OctoPrint configuration file
- A frame of your webcam
- The time when the print starts
IMPORTANT: the plugin only collects data when a print is running!
Why is there a MIT license in the project but you mentioned AGPLv3?
Some parts of the repository are licensed under different licenses and contain separate LICENSE files.
- Data Collector Plugin (AGPLv3) - borrows code from The Spaghetti Detective (which is a nice project, take a look. The authors also share their public dataset.)
- The dataset itself (not yet released) is licensed as ODbL
to be done, I'm sorry
- Publish the project on GitHub
- Publish the API
- Package the plugin and publish it as OctoPrint plugin
- Get this repository somehow spreaded into the world (This checkbox will never be checked)
- Find more contributors (Same applies here)
- Publish the first dataset
- Improve the Roadmap...
- Dennis
- 😲 your name can be here!
- 💤 it is quite empty here 💤
- Dennis
- 😲 your name can be here!
- 💤 it is quite empty here 💤
If you really read the entire page till here (and I admit it is pretty long and boring) it must be at least an interesting project so feel free to drop a 🌟 star or follow the project!
(PS: I just like GIFS and it is hopyfully eye-catching enough to make you read the above line)