Worked example: a re-labeled dataset release
示例
A team captures a warehouse robot episode on 12 March. The raw record includes camera frames, joint readings, and gripper force. The site agreement covers internal research and resale to third parties. The team annotates the episode, then six weeks later re-annotates it with a corrected labeling scheme. They want to sell the dataset. What should the provenance and versioning records contain, and how should the re-label be handled?
演示讲解
Step 1 — Record the capture conditions. The provenance record stores the capture date (12 March), the site, the task, and the sensor configuration (camera, joint encoders, gripper force). This is the first link in the chain and it cannot be reconstructed later.
Step 2 — Record the consent scope. The site agreement is logged with its scope: internal research and third-party resale. This is what makes the dataset sellable at all; without it, the capture is not licensed for resale regardless of data quality.
Step 3 — Record the raw source. The unmodified sensor stream is stored and given a stable identifier, so any later processing can be checked against the original.
Step 4 — Record the first annotation as a processing step. The initial labeling scheme, its version, and the date are logged as a step applied to the raw record. This produces dataset version v1.
Step 5 — Handle the re-label as a new version, not an overwrite. The corrected labeling scheme is logged as a new processing step with its own version and date, and it is recorded as derived from v1. This produces v2. The v1 files are kept, not replaced.
Step 6 — Name the release. The delivered dataset is identified as v2, with a pointer back to v1 and to the raw source. A buyer receiving v2 can trace every file back through the re-label to the original capture and its consent scope.