Dataset overview
ReFrame-Handheld contains 1,651 independent operation recordings across 17 acquisition batches and four tasks. The same brand and model of capture device is represented by three camera-extrinsic configurations. Configuration identifiers are used to associate each recording with its applicable intrinsics and camera-to-camera transforms.
Task coverage
| Task | Task ID | Recordings | Recorded minutes | Image streams |
|---|---|---|---|---|
| Carrot picking and placing | pick_carrot | 696 | 138.77 | 4 |
| Button pressing | press_button | 500 | 77.50 | 4 |
| Drawer opening | open_drawer | 255 | 41.69 | 4 |
| Cup hanging | hang_cup | 200 | 44.78 | 1 |
Recorded duration totals 302.75 minutes, approximately 5.046 hours. It is computed as the sum of each source recording's end-minus-start interval and includes recording margins. It is counted once per independent operation, not once per camera. Duplicate copies and derived visual versions do not increase the count.
Observation streams
| Camera role | Recordings | Frames | Resolution | Captured channels |
|---|---|---|---|---|
| main_rgb | 1651 | 991,958 | 1280×1280 | RGB |
| left_fisheye | 1451 | 464,362 | 640×480 | Grayscale |
| right_fisheye | 1451 | 464,345 | 640×480 | Grayscale |
| usb_reference | 1451 | 461,189 | 1920×1080 | RGB |
The exported set contains 6,004 MP4 files and 2,381,854 image frames. It retains 9,078,305 SLAM pose samples and 3,631,798 gripper samples at their original sampling rates. Episode payload totals approximately 67.25 GB, excluding repository-level files and the optional starter ZIP.
Main-camera record-time rates vary by batch, approximately 48–60 Hz; left/right cameras are approximately 30 Hz, and USB is typically approximately 29.8 Hz. These rates describe message timing, not independent hardware-frequency or exposure-synchronization measurements. Per-episode timestamps and machine-readable statistics contain the detailed information.
What users can study
- Inspect differences between real camera views of handheld operations.
- Develop and examine geometric reprojection or viewpoint-conversion methods using the applicable calibrated relationships.
- Compare transformed images with available real reference-camera observations, while accounting for timing and calibration limitations.
- Explore visual reuse of handheld demonstrations with retained motion and gripper records.
Reference images correspond only to the cameras that actually recorded them. The 200 main-only recordings do not have auxiliary-image ground truth. The resource does not supply arbitrary virtual target views, dense ground-truth depth, robot TCP mappings or policy-training actions. It is not a predefined benchmark leaderboard.
Task outcome, pose review, stream availability and time semantics are distinct metadata fields. The retained recordings are not uniformly labelled successful. No official train/test split is defined; any later split should consider acquisition batches and repeated scenes rather than treating task folders as evaluation splits.
Eight ColorChecker recordings are reserved as an optional separate attachment. They are not part of the operation count and have not yet been exported into this packaged format.