Hive: Explanation
Reviewing samples
ExplanationWho the reviewers are, how a reviewer works through the queue, what they check, and what it means for you if you own a machine.
AI-generated first draft. Written from Hive's and the machine UI's source and from community discussion, not from clicking through the flow start to finish. Correct it as you use it.
Samples are camera pictures your machine uploads to Hive while it runs. Each one has outlines drawn around the pieces in it. If that is new to you, How Hive works explains what a sample is first.
A program draws the outlines, so people check them. Those people are called reviewers. This page explains who they are and how they work.
What a reviewer is
A reviewer is a person who checks samples on Hive. For each sample, they vote accept or reject on the outlines.
- They see every machine's samples. A normal Hive account sees only the samples from its own machines.
- Several of them vote on each sample. One reviewer's vote never decides a sample alone.
- The Basically team gives out the role. Most people who use Hive are not reviewers, and you do not need to be one to build profiles or sort. If you want to help, ask in the
#sample-reviewchannel on Discord.
How a reviewer works
Reviewers work through a queue. Hive serves one sample at a time.
- Open the review screen. Hive shows the first sample in the queue, with its outlines drawn on the picture.
- Judge the outlines. Is each piece outlined correctly? The next section says what to look for.
- Vote. Accept, reject or skip. Use the buttons, or the keyboard: up accepts, down rejects, right skips and left goes back to the last sample.
- Move on. Hive serves the next sample. A reviewer keeps going until the queue is empty or they stop.
The queue only gives out fresh work. It hides samples the reviewer already voted on and samples that already have 3 votes. It also hides samples whose outlines have not been checked yet, and badly exposed shots. Each reviewer gets a different order, so two reviewers do not vote on the same sample at the same moment.
When a reviewer is unsure, they skip. Skipping costs nothing. A wrong vote can send a sample into conflict.
What a reviewer checks
- Accept when every piece on the target c-channel has its own tight outline.
- Accept when the channel is empty and there are no outlines.
- Reject when an outline is almost right but not tight.
- Reject when two touching pieces share one outline.
How votes become a result
A sample needs 3 votes. If all 3 accept, it is Accepted. If all 3 reject, it is Rejected. If they disagree, it is a Conflict. A conflict stays stuck until a reviewer changes a vote.
Reviewers find conflicts on the Samples page. They filter the status to Conflict and look at the outlines again before they vote.
Accepted samples are what a new piece-finding model is trained from. That is the reason the votes matter.
If you own a machine
You do not have to do anything. If your machine uploads samples, reviewers check them for you.
You can see your own machine's samples, and their status, under Samples in Hive. The status tells you how far a sample is: Unreviewed, In review, Accepted, Rejected or Conflict.
Voting is limited to reviewers. An account without the role cannot submit votes.
Records page or Hive
These are two different jobs, not the same job in two places.
- Records page, in your machine's own UI. You check pieces your machine actually sorted. Confirm or reject the part it named, and correct the colour. The answer goes back to the part-recognition service as feedback on that one result. Only pieces with a recognition result can be corrected. It does not count toward the 3 votes on Hive.
- Hive review. A reviewer judges the outlines on camera samples from the channels, the classification chamber and so on. Votes here decide whether a sample goes into the training set. That set trains the model that finds pieces in the camera picture.
If your aim is to correct what your own machine got wrong on a piece it sorted, use the Records page.
If the review queue is empty
This is for reviewers. Work down this list.
- Your machine and others. New samples appear only while a machine is running with sampling switched on. If machines are offline or not uploading, nothing new arrives.
- The filters. The default queue hides four kinds of sample: ones you already voted on, ones with 3 votes, ones whose outlines have not been checked yet, and badly exposed shots. The Unreviewed and Conflict status filters on Samples can still show work when the main queue looks empty.
- Your role. Without the reviewer role you see only your own machines' samples and cannot vote. Ask in
#sample-review.