Tasks
Details
Labels
Type limits which tool can use the label (any = boxes, polygons, lines, points). Skeleton labels need joints; drag the joints in the preview. Action labels (video) are things the person does, like "Cut bell pepper": annotators mark where each one starts and ends on a timeline under the video. Attributes are extra questions per object (colour, occluded…). "Changes" means the value can change from frame to frame in a video track.
Pictures / video
Choose JPG/PNG pictures or a whole folder. They go straight to your Backblaze bucket. Big folders are cut into batches: each batch becomes its own task (same labels and instructions), e.g. "… · batch 1", "… · batch 2". Pictures already uploaded are skipped, so if the upload stops, just choose the same folder again to continue.
One video per task (MP4, H.264 plays everywhere). For several videos, create one task per video.
🤖 AI pre-labels
Upload the …-prelabel.json from the Colab Auto Pre-label notebook (same video, same label JSON as this task). Every empty job of this task then opens with the AI boxes and draft actions already in place; annotators fix them and save. Jobs that already have work are not changed. Never upload the preview MP4 as the task video.
Answer key (training)
Upload the correct masks (black/white pictures, white = the object) with the same file names as the pictures above, e.g. from the BDAPPV mask folder. Annotators never see them. Reviewers get a 🎯 button in the editor, and each job gets an accuracy % (overlap with the answer).
Jobs
People on this task
Used when "Who can take jobs" is "Only people added to this task" (step 1). Filter, tick people, then Add to task.
Training leaderboard
Score = pictures/frames in accepted jobs; jobs accepted on the first try count ×1.5; jobs with an answer key count × their accuracy. Ties: higher first-try %, then faster. Workers see this on their dashboard (first name + last initial). Tick "Training task" in step 1 to include a task.
Export for the client
Picture file names are kept, so the client can match them to the originals. For video, frames are named frame_000123.jpg (frame number in the video). Video frames themselves are not included: the client uses the original video.