Frequently Asked Questions¶

Why are you using Tator and not VARS or FathomNet?¶

VARS and FathomNet are great for what they do, but our projects need some specific features they don't offer—like flexible metadata, versioning, and a web UI for annotation. We also need to be able to run everything on-site where our data lives, share result links easily, and export data whenever we need to. We’re always exploring new tech to meet our needs, including our recent work with Voxel51.

How much data have these tools been used on?¶

What kind of projects are these tools useful for?¶

If your project involves classifying or detecting objects in just about any kind of media, these tools can help. As long as the data can be represented as an image—like sound spectrograms, drone photos, time-lapse shots, or video frames—you’re good to go.

Do I need to label all of my data to train a model?¶

Short answer: Nope!

Long answer: You can start by labeling a small chunk of your data to train an initial model. Once that’s running, you can use it to help label the rest. We’ve found that doing this twice is usually enough to get most datasets labeled, especially if you use our sdcat clustering, our vector database, or pseudo-labeling to speed things up.

For a good classification model, 100 examples is a decent start, but 300 is a better target. More is usually better, but not always—having thousands of examples can sometimes just slow down your training for no real gain. For detection models, aim for about 1,000 frames per class, though you can sometimes get away with less. Just remember that detection models take longer to prep because you have to label everything in the frame. We’ve found that it’s often faster to start with a simple binary classification (like "target" vs "background") and then refine things from there.

Flexible Taxonomies

Don't let rigid taxonomies slow down your ML development. See Flexible Taxonomies in Machine Learning for guidance on balancing scientific taxonomies with practical ML labeling strategies.

I want to use these tools to label my data. Does it work on video?¶

Yes! You can upload your video to our Mantis server (which hosts Tator). The video will be transcoded so it's optimized for web viewing and works well even on slower connections. This can take a little while depending on the video size and how busy the server is, but once it’s done, you’ll be ready to start annotating in Tator.

Here’s a quick example of how to upload a video using the mbari-aidata package:

Install it:

pip install mbari-aidata

Then, register your video with something like this:

aidata upload \
    --config https://docs.mbari.org/internal/ai/projects/config/config_uav.yml \
    --base-path $PWD \
    --version Baseline \
    --token <TATOR_TOKEN> \
    --video input.mp4 \
    --upload

You can find your Tator token in the "API token" link after you log in.

How is the data stored?¶

We typically store metadata about your images, the images themselves, and their feature vectors. Here’s how it works:

Tator (PostgreSQL): Stores all the metadata for your images (location, size, depth, etc.).
Vector Database: Stores feature vectors—numerical summaries of your data. For an image, this might be a set of 768 numbers that represent its shapes and textures.

Is Tator the same tool FathomNet uses?¶

Yes and no. They use the same Tator database, but the user interface is different.

Updated: 2026-02-08