DataMachine / README.md
shchon11's picture
Update README.md
4b8dfdc verified
|
Raw History Blame
20.3 kB
# DM nuScenes — release `1010`
Driving data from the DM test vehicle (an Ouster LiDAR, 14 RGB cameras and 2 thermal cameras, GNSS/INS) in the
[nuScenes](https://www.nuscenes.org/) `v1.0-trainval` format: load it with the `nuscenes-devkit` as is.
**682 scenes** of 20 s (3.79 h) from **125 recordings** · 27,280 key-frame samples ·
6,512,042 sensor frames · 1665 GB to download in 268 sensor tars + 1 table tar
(1774 GB unpacked).
No 3D box annotations (`sample_annotation` is empty).
This release holds the recordings delivered so far (amsa, chuncheon, gangbyeonbungno, gangnam, gapyeong, geondae, godeok, gosokdoro, jongno, magok, namyangju, ollimpikdaero, seongsu, wangsimni, yanggu); the tables list exactly the scenes
whose files are in it. More recordings are added by later releases.
## Datasets in this repository
| folder | dataset | scenes | recordings | GB |
|---|---|---|---|---|
| repository root | dm_nuscenes | 682 | 125 | 1665 | ← this card
| `dm_nuscenes_2/` | dm_nuscenes_2 | 103 | 6 | 292 |
Each is a complete nuScenes `v1.0-trainval` dataroot of its own. A dataset holds at most the 685 official train scene
names, so the next one starts the list again (`scene-0001`, ...): their scene names clash. Put each together into
its own folder and open each with its own `NuScenes(dataroot=...)`. The same release (branch) holds all of them.
## Download and use
1. Log in if the dataset is gated or private: `hf auth login`.
2. Download this release and put it together (`assemble.py` checks every tar's sha256 and extracts it):
```bash
pip install -U huggingface_hub
hf download YewonSong/DataMachine --repo-type dataset --revision 1010 --local-dir dm_tars
python dm_tars/assemble.py --tars dm_tars --out dm_nuscenes # -> dm_nuscenes/ and dm_nuscenes_2/
```
One dataset only: the root one with `--exclude "dm_nuscenes_2/*"` on `hf download`.
```python
from nuscenes.nuscenes import NuScenes
from nuscenes.can_bus.can_bus_api import NuScenesCanBus
nusc = NuScenes(version="v1.0-trainval", dataroot="dm_nuscenes")
can = NuScenesCanBus(dataroot="dm_nuscenes") # pose, ms_imu, meta per scene
```
**Some recordings only**: download the tables and the recordings you want, and let `assemble.py` skip the rest
(the tables still list every scene of the release; the files of the others are just absent):
```bash
hf download YewonSong/DataMachine --repo-type dataset --revision 1010 --local-dir dm_tars \
--include "manifest.json" --include "assemble.py" --include "meta/*" --include "sensors/rec_20260929_162450/*"
python dm_tars/assemble.py --tars dm_tars --out dm_nuscenes --allow-missing
```
**Updating**: Coming from `1007`: run the two commands above in the same directories — only the tars new in `1010` are downloaded and extracted. Nothing already extracted is changed;
the tables are replaced. Add `--delete-tars` to `assemble.py` to remove each tar after extracting it (then
`hf download` fetches it again next time).
## Releases and versioning
Each release is a branch of this repository (`--revision <name>`), a complete dataset on its own; a release
is never changed after it is published. Later releases only **add** data:
- `sensors/<recording>/<recording>.partNN.tar.gz` — the sensor files (`samples/`, `sweeps/`) of a group of whole
scenes of one recording, at most 10 GB unpacked (gzip: the LiDAR files shrink by about a third, the images not).
Once published, a tar never changes and is carried into every later release (stored once).
- `meta/DM_meta.tar.gz` — the tables (`v1.0-trainval/`) of the scenes released so far, `can_bus/`, `maps/` and the
per-recording `*.import.json` (how each recording was converted; the driver's and passenger's names are removed).
The only file that changes between releases.
- `manifest.json` — every tar with its recording, scene tokens, files, bytes and sha256, and the release history.
- Scene tokens never change. Scene names (`scene-0001`, …) are the nuScenes train split names in the order the
recordings were converted, so they are not contiguous inside one release; the names of released scenes do not
change. All scenes are in the `train` split.
| release | date | locations added | recordings added | scenes added | scenes in total | parser commit |
|---|---|---|---|---|---|---|
| `1002` | 2026-10-03 | amsa | 4 | 80 | 80 | `56ad0fe` (+local changes) |
| `1003` | 2026-10-06 | gangbyeonbungno, godeok | 13 | 121 | 201 | `56ad0fe` (+local changes) |
| `1007` | 2026-10-08 | chuncheon, gangbyeonbungno, gangnam, gapyeong, geondae, gosokdoro, namyangju, seongsu, wangsimni, yanggu | 88 | 436 | 637 | `56ad0fe` (+local changes) |
| `1010` | 2026-10-10 | jongno, magok, ollimpikdaero, wangsimni | 20 | 45 | 682 | `56ad0fe` (+local changes) |
## Contents
| location | dates | recordings | scenes | time | GB |
|---|---|---|---|---|---|
| amsa | 2026-09-29 | 4 | 80 | 27 min | 202 |
| chuncheon | 2026-10-01 | 4 | 10 | 3 min | 14 |
| gangbyeonbungno | 2026-09-29, 2026-10-03 | 7 | 7 | 2 min | 13 |
| gangnam | 2026-10-02 | 9 | 154 | 51 min | 413 |
| gapyeong | 2026-10-01 | 2 | 3 | 1 min | 6 |
| geondae | 2026-10-01 | 15 | 80 | 27 min | 208 |
| godeok | 2026-09-29 | 7 | 115 | 38 min | 300 |
| gosokdoro | 2026-10-03 | 3 | 4 | 1 min | 4 |
| jongno | 2026-10-06 | 15 | 36 | 12 min | 84 |
| magok | 2026-10-07 | 2 | 2 | 1 min | 6 |
| namyangju | 2026-10-03 | 34 | 64 | 21 min | 143 |
| ollimpikdaero | 2026-10-07 | 2 | 4 | 1 min | 10 |
| seongsu | 2026-10-01 | 9 | 58 | 19 min | 143 |
| wangsimni | 2026-10-01, 2026-10-06 | 5 | 29 | 10 min | 76 |
| yanggu | 2026-10-03 | 7 | 36 | 12 min | 42 |
| channel | sensor | data | rate | frames |
|---|---|---|---|---|
| `CAM_FRONT_1` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_2` | camera | 1920×1200 JPG | 29 Hz | 399,648 |
| `CAM_FRONT_3` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_4` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_5` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_6` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_7` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_8` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_FRONT_9` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_REAR_LEFT` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_REAR_RIGHT` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_SIDE_LEFT` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_SIDE_RIGHT` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `CAM_THERMAL_LEFT` | thermal camera | 640×480 PNG | 29 Hz | 391,464 |
| `CAM_THERMAL_RIGHT` | thermal camera | 640×480 PNG | 29 Hz | 391,808 |
| `CAM_TOP` | camera | 1920×1200 JPG | 29 Hz | 399,650 |
| `LIDAR_TOP` | lidar | point cloud (.pcd.bin) | 10 Hz | 133,672 |
Key frames (`samples/`) are 2 Hz, synchronised across the sensors as in nuScenes; every other frame is a sweep
(`sweeps/`). LiDAR files hold five float32 per point (x, y, z, intensity, ring), as in nuScenes. The camera images
are stored as recorded (not rectified): `calibrated_sensor` holds `camera_intrinsic` and, for the cameras,
`distortion_model` and `camera_distortion` (lens model and coefficients), plus the extrinsics, one set per recording.
## Coordinate frames and the quality of the pose
`ego_pose` is in the global frame of the recording's **location**, as in nuScenes: metres, x east, y north, z up,
on the tangent plane at that location's origin (height 0 on the WGS84 ellipsoid). `log.json` names it in the extra
key `global_frame` (`enu@37.550000,127.140000,0.000`) and `location` (`korea-amsa`, `korea-godeok`, ...). Each
location has its own origin: positions of different locations are not in one plane.
**The pose is poor.** It comes from the vehicle's GNSS/INS (`/gps/fix`, `/imu/data`, `/gps/vel`; 100 Hz), and the
receiver often has no GNSS solution (the pose is then the INS coasting), so absolute positions can drift or jump
and the speed can be off. Per-scene figures are in `<recording>.import.json` (`pos_type`, `gnss_solution_by_scene`,
`position_sigma_m_median`). Use the poses for what they are good for (ego orientation, short-term motion), not
as ground truth; the IMU (`can_bus/<scene>_ms_imu.json`) and the LiDAR are more reliable than the position.
## Recordings
| recording | location | date | scenes | tars | GB | first release |
|---|---|---|---|---|---|---|
| rec_20260929_144612 | godeok | 2026-09-29 | 23 | 7 | 54.9 | `1003` |
| rec_20260929_150138 | godeok | 2026-09-29 | 14 | 5 | 36.0 | `1003` |
| rec_20260929_151129 | godeok | 2026-09-29 | 21 | 7 | 54.7 | `1003` |
| rec_20260929_152855-카메라2죽음 | godeok | 2026-09-29 | 7 | 3 | 19.1 | `1003` |
| rec_20260929_154234 | godeok | 2026-09-29 | 25 | 8 | 68.4 | `1003` |
| rec_20260929_155714 | godeok | 2026-09-29 | 14 | 5 | 38.3 | `1003` |
| rec_20260929_160819 | godeok | 2026-09-29 | 11 | 4 | 28.7 | `1003` |
| rec_20260929_162450 | amsa | 2026-09-29 | 25 | 9 | 64.8 | `1002` |
| rec_20260929_163610 | amsa | 2026-09-29 | 19 | 6 | 45.4 | `1002` |
| rec_20260929_164535 | amsa | 2026-09-29 | 16 | 6 | 40.4 | `1002` |
| rec_20260929_165424 | amsa | 2026-09-29 | 20 | 7 | 51.9 | `1002` |
| clip_20260929_171328 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 1.7 | `1003` |
| clip_20260929_171941 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 2.1 | `1003` |
| clip_20260929_172040 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 2.0 | `1003` |
| clip_20260929_172123 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 1.9 | `1003` |
| clip_20260929_172245 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 1.9 | `1003` |
| clip_20260929_172502 | gangbyeonbungno | 2026-09-29 | 1 | 1 | 2.0 | `1003` |
| rec_20261001_193910 | seongsu | 2026-10-01 | 10 | 4 | 27.2 | `1007` |
| rec_20261001_194603 | seongsu | 2026-10-01 | 14 | 5 | 37.0 | `1007` |
| rec_20261001_195123 | seongsu | 2026-10-01 | 3 | 1 | 7.0 | `1007` |
| rec_20261001_195257 | seongsu | 2026-10-01 | 2 | 1 | 4.3 | `1007` |
| rec_20261001_195915 | geondae | 2026-10-01 | 2 | 1 | 5.1 | `1007` |
| rec_20261001_200100 | geondae | 2026-10-01 | 12 | 4 | 34.4 | `1007` |
| rec_20261001_200556 | geondae | 2026-10-01 | 12 | 4 | 34.5 | `1007` |
| rec_20261001_201052 | geondae | 2026-10-01 | 2 | 1 | 4.8 | `1007` |
| rec_20261001_201333 | geondae | 2026-10-01 | 10 | 3 | 21.9 | `1007` |
| rec_20261001_202047 | geondae | 2026-10-01 | 1 | 1 | 2.4 | `1007` |
| rec_20261001_202153 | geondae | 2026-10-01 | 7 | 3 | 18.7 | `1007` |
| rec_20261001_202525 | geondae | 2026-10-01 | 1 | 1 | 2.1 | `1007` |
| rec_20261001_202626 | geondae | 2026-10-01 | 4 | 2 | 10.8 | `1007` |
| rec_20261001_202817 | geondae | 2026-10-01 | 4 | 1 | 9.2 | `1007` |
| rec_20261001_203119 | geondae | 2026-10-01 | 1 | 1 | 2.4 | `1007` |
| rec_20261001_203252 | geondae | 2026-10-01 | 11 | 4 | 31.9 | `1007` |
| rec_20261001_203908 | geondae | 2026-10-01 | 2 | 1 | 5.1 | `1007` |
| rec_20261001_203959 | geondae | 2026-10-01 | 7 | 2 | 14.1 | `1007` |
| rec_20261001_204300 | geondae | 2026-10-01 | 4 | 2 | 10.8 | `1007` |
| rec_20261001_204850 | seongsu | 2026-10-01 | 3 | 1 | 6.3 | `1007` |
| rec_20261001_205137 | seongsu | 2026-10-01 | 13 | 4 | 30.8 | `1007` |
| rec_20261001_210116 | seongsu | 2026-10-01 | 5 | 2 | 11.6 | `1007` |
| rec_20261001_210622 | seongsu | 2026-10-01 | 5 | 2 | 12.3 | `1007` |
| rec_20261001_211024 | seongsu | 2026-10-01 | 3 | 1 | 6.7 | `1007` |
| rec_20261001_211509 | wangsimni | 2026-10-01 | 2 | 1 | 4.1 | `1007` |
| rec_20261001_211716 | wangsimni | 2026-10-01 | 3 | 1 | 8.6 | `1007` |
| rec_20261001_211842 | wangsimni | 2026-10-01 | 13 | 5 | 34.1 | `1007` |
| rec_20261001_212356 | wangsimni | 2026-10-01 | 8 | 3 | 21.7 | `1007` |
| rec_20261002_063935 | chuncheon | 2026-10-01 | 3 | 1 | 3.8 | `1007` |
| rec_20261002_065008 | chuncheon | 2026-10-01 | 2 | 1 | 2.7 | `1007` |
| rec_20261002_070000 | chuncheon | 2026-10-01 | 1 | 1 | 1.4 | `1007` |
| rec_20261002_070515 | chuncheon | 2026-10-01 | 4 | 1 | 5.9 | `1007` |
| rec_20261002_075439 | gapyeong | 2026-10-01 | 1 | 1 | 1.8 | `1007` |
| rec_20261002_080046 | gapyeong | 2026-10-01 | 2 | 1 | 4.2 | `1007` |
| rec_20261002_122400 | gangnam | 2026-10-02 | 31 | 11 | 79.3 | `1007` |
| rec_20261002_124459 | gangnam | 2026-10-02 | 7 | 3 | 18.1 | `1007` |
| rec_20261002_130230 | gangnam | 2026-10-02 | 29 | 9 | 75.6 | `1007` |
| rec_20261002_131637 | gangnam | 2026-10-02 | 7 | 3 | 20.4 | `1007` |
| rec_20261002_132644 | gangnam | 2026-10-02 | 32 | 11 | 91.0 | `1007` |
| rec_20261002_134111 | gangnam | 2026-10-02 | 10 | 4 | 28.6 | `1007` |
| rec_20261002_134922 | gangnam | 2026-10-02 | 10 | 4 | 25.2 | `1007` |
| rec_20261002_144625 | gangnam | 2026-10-02 | 3 | 1 | 7.5 | `1007` |
| rec_20261002_144911 | gangnam | 2026-10-02 | 25 | 9 | 67.0 | `1007` |
| clip_20261003_155736 | namyangju | 2026-10-03 | 1 | 1 | 2.3 | `1007` |
| clip_20261003_155916 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_160515 | namyangju | 2026-10-03 | 3 | 1 | 6.6 | `1007` |
| rec_20261003_161221 | namyangju | 2026-10-03 | 1 | 1 | 2.2 | `1007` |
| rec_20261003_161922 | namyangju | 2026-10-03 | 1 | 1 | 2.2 | `1007` |
| rec_20261003_162600 | namyangju | 2026-10-03 | 2 | 1 | 4.4 | `1007` |
| clip_20261003_162808 | namyangju | 2026-10-03 | 1 | 1 | 2.3 | `1007` |
| clip_20261003_163137 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_164207 | namyangju | 2026-10-03 | 4 | 1 | 9.0 | `1007` |
| rec_20261003_164412 | namyangju | 2026-10-03 | 4 | 2 | 9.8 | `1007` |
| rec_20261003_164612 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_164748 | namyangju | 2026-10-03 | 1 | 1 | 2.7 | `1007` |
| rec_20261003_164852 | namyangju | 2026-10-03 | 1 | 1 | 2.5 | `1007` |
| rec_20261003_165057 | namyangju | 2026-10-03 | 2 | 1 | 4.7 | `1007` |
| rec_20261003_165205 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_165423 | namyangju | 2026-10-03 | 1 | 1 | 2.3 | `1007` |
| rec_20261003_165623 | namyangju | 2026-10-03 | 2 | 1 | 4.9 | `1007` |
| rec_20261003_165830 | namyangju | 2026-10-03 | 2 | 1 | 5.1 | `1007` |
| rec_20261003_185214 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_185348 | namyangju | 2026-10-03 | 3 | 1 | 6.2 | `1007` |
| rec_20261003_185713 | namyangju | 2026-10-03 | 4 | 1 | 7.3 | `1007` |
| clip_20261003_190137 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_190352 | namyangju | 2026-10-03 | 1 | 1 | 2.7 | `1007` |
| rec_20261003_190449 | namyangju | 2026-10-03 | 3 | 1 | 8.3 | `1007` |
| rec_20261003_190606 | namyangju | 2026-10-03 | 1 | 1 | 2.5 | `1007` |
| rec_20261003_190744 | namyangju | 2026-10-03 | 1 | 1 | 2.3 | `1007` |
| rec_20261003_191016 | namyangju | 2026-10-03 | 9 | 3 | 19.1 | `1007` |
| rec_20261003_191425 | namyangju | 2026-10-03 | 2 | 1 | 4.2 | `1007` |
| rec_20261003_191521 | namyangju | 2026-10-03 | 1 | 1 | 2.4 | `1007` |
| rec_20261003_191655 | namyangju | 2026-10-03 | 1 | 1 | 2.0 | `1007` |
| rec_20261003_191738 | namyangju | 2026-10-03 | 2 | 1 | 3.6 | `1007` |
| rec_20261003_191919 | namyangju | 2026-10-03 | 1 | 1 | 1.7 | `1007` |
| rec_20261003_192014 | namyangju | 2026-10-03 | 2 | 1 | 3.5 | `1007` |
| rec_20261003_192237 | namyangju | 2026-10-03 | 1 | 1 | 2.0 | `1007` |
| clip_20261003_193542 | gangbyeonbungno | 2026-10-03 | 1 | 1 | 1.8 | `1007` |
| rec_20261004_033655 | gosokdoro | 2026-10-03 | 1 | 1 | 1.2 | `1007` |
| rec_20261004_035019 | gosokdoro | 2026-10-03 | 1 | 1 | 1.1 | `1007` |
| rec_20261004_051712 | gosokdoro | 2026-10-03 | 2 | 1 | 2.1 | `1007` |
| rec_20261004_053533 | yanggu | 2026-10-03 | 7 | 1 | 7.1 | `1007` |
| rec_20261004_055315 | yanggu | 2026-10-03 | 12 | 2 | 14.2 | `1007` |
| rec_20261004_060022 | yanggu | 2026-10-03 | 6 | 1 | 6.4 | `1007` |
| rec_20261004_062938 | yanggu | 2026-10-03 | 2 | 1 | 2.8 | `1007` |
| rec_20261004_064254 | yanggu | 2026-10-03 | 5 | 1 | 6.4 | `1007` |
| rec_20261004_064745 | yanggu | 2026-10-03 | 3 | 1 | 3.8 | `1007` |
| rec_20261004_065138 | yanggu | 2026-10-03 | 1 | 1 | 1.2 | `1007` |
| rec_20261006_172715 | wangsimni | 2026-10-06 | 3 | 1 | 7.9 | `1010` |
| rec_20261006_173126 | jongno | 2026-10-06 | 2 | 1 | 4.9 | `1010` |
| rec_20261006_173256 | jongno | 2026-10-06 | 3 | 1 | 6.9 | `1010` |
| rec_20261006_173848 | jongno | 2026-10-06 | 1 | 1 | 2.4 | `1010` |
| rec_20261006_174000 | jongno | 2026-10-06 | 3 | 1 | 8.3 | `1010` |
| rec_20261006_174315 | jongno | 2026-10-06 | 1 | 1 | 2.6 | `1010` |
| rec_20261006_174551 | jongno | 2026-10-06 | 1 | 1 | 1.9 | `1010` |
| rec_20261006_174838 | jongno | 2026-10-06 | 5 | 2 | 12.5 | `1010` |
| rec_20261006_175735 | jongno | 2026-10-06 | 1 | 1 | 2.8 | `1010` |
| rec_20261006_175946 | jongno | 2026-10-06 | 2 | 1 | 4.6 | `1010` |
| rec_20261006_180129 | jongno | 2026-10-06 | 7 | 3 | 16.9 | `1010` |
| rec_20261006_180519 | jongno | 2026-10-06 | 1 | 1 | 2.1 | `1010` |
| rec_20261006_180644 | jongno | 2026-10-06 | 2 | 1 | 4.4 | `1010` |
| rec_20261006_180802 | jongno | 2026-10-06 | 4 | 1 | 8.0 | `1010` |
| rec_20261006_181026 | jongno | 2026-10-06 | 2 | 1 | 4.1 | `1010` |
| clip_20261006_181249 | jongno | 2026-10-06 | 1 | 1 | 1.4 | `1010` |
| rec_20261007_110904 | magok | 2026-10-07 | 1 | 1 | 2.9 | `1010` |
| rec_20261007_133223 | magok | 2026-10-07 | 1 | 1 | 2.8 | `1010` |
| rec_20261007_140204 | ollimpikdaero | 2026-10-07 | 1 | 1 | 2.3 | `1010` |
| rec_20261007_140543 | ollimpikdaero | 2026-10-07 | 3 | 1 | 8.1 | `1010` |
## Maintainers: making the next release
Releases are made with `parser/scripts/hf_release.py` of the DM parser (`IRCV-DM-Ops`; the commit of each release is in
the table above):
```bash
# new recordings: parse them, curate them in the DM Parser app, press 최종 확정, then
python parser/scripts/hf_release.py --branch <new> --from 1010 --location <place> --token-file <hf token file> # plan
python parser/scripts/hf_release.py --branch <new> --from 1010 --location <place> --token-file <hf token file> --upload # build + upload
```
It refuses to upload unless the scenes it adds are locked and the dataset passes the checks (table references,
devkit, CAN bus), and unless every tar already published still matches the dataset. Only the new recordings'
tars are built and uploaded (one commit each, resumable); the tables, manifest, this card and `assemble.py`
go last. First release of this series: `1002`.
씬 이름이 모자라서 나눴습니다. 한 폴더에는 공식 train 이름 685개까지만 들어갈 수 있고, DM은 그걸 넘었습니다.
이름이 왜 제한되는가
nuScenes devkit은 씬 이름으로 train/val을 구분합니다. scene-0001 같은 이름이 공식 분할 목록에 있어야 합니다.
CAN bus API도 정해진 이름만 받습니다.
그래서 DM 데이터를 devkit으로 "그대로" 읽히게 하려고, 처음부터 공식 train 이름을 순서대로 붙였습니다.
그 이름이 685개뿐이라, 686번째 씬부터는 붙일 이름이 없습니다.
그래서 두 번째 폴더를 만든 것
dm_nuscenes_2는 이름을 scene-0001부터 다시 씁니다.
두 데이터셋에 같은 이름의 씬이 있게 되므로, 한 폴더에 섞으면 충돌합니다. 그래서 각각 따로 여는 별개의 dataroot로 뒀습니다.
한 폴더로 합치려면 다음 중 하나를 감수해야 합니다.
방법 대가
val 이름(150개)까지 써서 850개로 늘림 일부 씬이 devkit에서 val로 분류됨. 850개를 넘으면 또 나눠야 함
공식 목록에 없는 이름(scene-2001 등)을 새로 붙임 devkit의 분할 구분과 CAN bus API가 그 씬에서 동작하지 않음
지금처럼 폴더를 나눔 사용자가 dataroot를 두 번 열어야 함
지금 방식은 devkit 호환을 온전히 지키는 대신 폴더가 나뉘는 쪽을 택한 것입니다. 다만 이미 HF에 scene-XXXX 이름으로 올라간 씬이 785개라, 방식을 바꾸면 이미 올린 씬의 이름이 바뀝니다. 업체가 받아 간 뒤라면 혼란이 생길 수 있습니다.
이 구조는 처음 685개를 넘겼을 때(10/5) 제가 그렇게 정한 것이고, 따로 여쭤보고 정한 건 아니었습니다. 학습할 때 train/val을 직접 나눌 계획이라 devkit의 분할 구분이 필요 없다면, 한 폴더로 합치는 쪽이 쓰기에는 더 편합니다. 원하시면 그 방향으로 바꾸는 방법과 영향을 정리해 드리겠습니다.