Add public dataset card and privacy notes
Browse files
README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
pretty_name: Intersection Maneuver v32
|
| 3 |
+
license: cc-by-4.0
|
| 4 |
+
language:
|
| 5 |
+
- zh
|
| 6 |
+
- en
|
| 7 |
+
task_categories:
|
| 8 |
+
- time-series-classification
|
| 9 |
+
tags:
|
| 10 |
+
- cycling
|
| 11 |
+
- pose-estimation
|
| 12 |
+
- intersection
|
| 13 |
+
- gesture-recognition
|
| 14 |
+
- maneuver-prediction
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Intersection Maneuver v32
|
| 18 |
+
|
| 19 |
+
This public research release contains anonymized 2D Skeleton CSV sequences, Gesture and
|
| 20 |
+
Maneuver labels, participant-disjoint split manifests, selected model checkpoints, and
|
| 21 |
+
sanitized aggregate evaluation results for intersection cyclist-intention classification.
|
| 22 |
+
|
| 23 |
+
## Current label release
|
| 24 |
+
|
| 25 |
+
Use `latest_labels_v32/` for the current labels. The label corrections were introduced in
|
| 26 |
+
v23, the official balanced participant-disjoint split was fixed in v31, and v32 changed
|
| 27 |
+
prediction calibration only.
|
| 28 |
+
|
| 29 |
+
- 599 anonymous Skeleton CSV files
|
| 30 |
+
- 566 labeled passages
|
| 31 |
+
- 12,569 Gesture windows
|
| 32 |
+
- 607 natural Maneuver windows
|
| 33 |
+
- 405 rows in the official v31 uniform 7:2:1 split
|
| 34 |
+
- classes: `Crossing`, `LeftTurn`, and `RightTurn`
|
| 35 |
+
|
| 36 |
+
## Models and evaluation
|
| 37 |
+
|
| 38 |
+
- `models/intersection_v31_official/`: official three-seed v31 baseline
|
| 39 |
+
- `models/intersection_v32_crossing_calibrated_candidate/`: calibration candidate pending
|
| 40 |
+
independent confirmation
|
| 41 |
+
- `evaluation/v31/` and `evaluation/v32/`: sanitized aggregate results, loss curves, confusion
|
| 42 |
+
matrices, and experiment summaries
|
| 43 |
+
- `v22/`: earlier participant-disjoint data, selected checkpoints, reports, and figures
|
| 44 |
+
|
| 45 |
+
## Privacy and exclusions
|
| 46 |
+
|
| 47 |
+
No raw RGB or Depth video is included. Participant identities, original filenames, session
|
| 48 |
+
identifiers, absolute timestamps, and local filesystem paths have been removed. Two private
|
| 49 |
+
identity-review videos and v22 per-sample prediction tables containing legacy source IDs are
|
| 50 |
+
intentionally excluded from this public release.
|
| 51 |
+
|
| 52 |
+
## Evaluation note
|
| 53 |
+
|
| 54 |
+
The official v31 Test contains 39 unique events, with 13 events per Maneuver class. The v32
|
| 55 |
+
Crossing calibration candidate was inspected alongside multiple alternatives on that Test and
|
| 56 |
+
therefore requires confirmation on new untouched data before promotion.
|