To answer your questions from what I know,
Q: Support Subscription:
Do I need to purchase a support, subscription to access technical assistance and customer support? I’d like to know the options available and the associated costs. Because I contact form via email and I didn’t get back any response.
A: No, anyone can receive support often times if the problem cant be solved via the community help page (here) employees reach out and enter the room directly via allowing access to your workspace to the employee, there is a support email however it may take some time to get back from those, however I can see you have the priority star which means you should be receiving help before those who do not have this feature.
Q: System Requirements:
Could you kindly recommend the minimum and recommended specifications for computer hardware to run that project efficiently on my local computer? I want to ensure that my system meets the necessary requirements for optimal performance.
A: Depending on how you’re using your object detection this varies. Google Collab with premium offers a hosted GPU with 16gb of DRAM, Roboflow also using credits allows you to train and work with models, also using Hosted GPU’s however if you want to do this all on your own machinery it will be costly, its recommended to go with hosted unless you have good GPU power, if you are curious of a good machine learning build, id recommend 64gb of DDR5 RAM, an RTX 3090 or greater, a intel i9-12900k or better, and plenty of power, say 1300W+. Again i’d recommend going hosted for this.
Q: Label Detection Issues(using object Detection):
In my last testing project, I encountered some difficulties with label detection. I’ve attached some images (Attachment 1) for your reference. Could you please provide guidance or solutions to improve label detection accuracy? Now, I’m using my friend’s account to ask that question.
A: I couldnt access the images, but (and im just guessing here) if you made a test model and it didnt preform so well, maybe the object detection etc, this is generally due to a poor dataset training for the model which is almost normal when first getting into making models. Simply put for good object detection you need plenty of data, usually thousand(s) of sample images that are well annotated.
Q: Long-Term Usage Preparation:
For medium-sized industry usage, what steps should I take to ensure the long-term reliability and stability of my software using your AI Object Detection? Are there any best practices or considerations I should be aware of?
A: I’m gonna pass this one off to staff, but the first thing I would suggest is build a solid model and then export it by downloading the weights on a separate server / computer. Somewhere safe where you have control over it.
Q: Camera Selection:
I have chosen to use the GoPro Hero 12 Black and the Logitech BRIO Ultra HD webcam for my project. Are these camera choices suitable for your software, or do you recommend specific brands or models for better compatibility and performance?
A: It mostly depends on the type of project you’re going for, id say the gopro is a great choice as it offers solid footage and great framerates, the brio is a good choice but not as good as the gopro. Things you’ll want to look for is stable performance at high resolutions which for most use-cases I think the gopro would do well at.
Q: Image Quantity for Classification and Detection:
For our project, we need to classify and detect a range of labels, from 13 to a maximum of 30. Could you provide guidance on how many images we should provide for classification and detection for these labels to ensure accurate results?
A: Okay, so this is a big project. Going off what I know about Microsoft COCO which uses:
- Over 200,000 labeled images.
- 1.5 million object instances.
- 80 object categories.
- 5 captions per image, leading to a total of 1 million captions.
Also important to take into account:
Small dataset: 1,000 images per class. This might work if you’re doing transfer learning or if the task is straightforward.
Medium dataset: 5,000 images per class. This provides a better representation, especially if there’s some variation within classes.
Large dataset: 10,000+ images per class. This is desirable for more complex tasks with a lot of intra-class variations.
So again it does matter about what you’re trying to do but i’d say a good rough estimate for a complex topic like this would be 30,000+ images.
Q: Transfer of Model and Roboflow:
In the event that I pass on this project to another user, what is the process for transferring the model and Roboflow setup to ensure a smooth transition? Your guidance on this matter would be greatly appreciated.
A: So the model weights you can give to anyone you choose once the model is done and ready to be shipped. However if you want to transfer the Workspace to someone you can invite a new user to the workspace and change their role before leaving if you so choose. You can also choose to make the model go public on Roboflow Universe if you wanted as well.