geo4d_evals / dynpose_eval /CAMERA_CONTROLS.md
HugMinjun's picture
Add files using upload-large-folder tool
c761871 verified
|
Raw History Blame Contribute Delete
1.79 kB
# DynPose synthetic camera controls
DynPose의 RGB와 ViPE OpenCV `c2w`는 관측 source trajectory다. Novel-camera
평가를 위해 각 source에 NeoVerse `CameraTrajectory.from_predefined` 기본값의
12개 relative instruction을 합성한다.
| trajectory | operation | relative endpoint translation |
|---|---|---|
| `pan_left` / `pan_right` | yaw ±15° | `[0, 0, 0]` |
| `tilt_up` / `tilt_down` | pitch ±15° | `[0, 0, 0]` |
| `move_left` / `move_right` | horizontal translation | `[∓0.1, 0, 0]` |
| `push_in` / `pull_out` | forward/backward translation | `[0, 0, ±0.1]` |
| `boom_up` / `boom_down` | vertical translation | `[0, ∓0.1, 0]` |
| `orbit_left` / `orbit_right` | 15° arc, radius 1.0 | `[∓0.258819, 0, 0.034074]` |
각 relative path는 source와 같은 83–256프레임으로 샘플링한다. 첫 relative
pose는 identity이고 마지막 pose는 표의 endpoint다. Rotation은 Slerp,
translation은 선형 보간한다. Absolute target control은 다음과 같다.
```text
target_c2w[t] = source_c2w[t] @ relative_c2w[t]
```
따라서 target frame 0은 관측 source frame 0과 정확히 같다. Distance 0.1과
orbit radius 1.0은 물리 미터가 아니라 scene별 ViPE reconstruction unit이다.
scene 사이의 절대 이동량으로 비교하지 않는다.
## Files
- Variable-length endpoint template: `trajectory_specs/<trajectory_id>.json`
- Absolute OpenCV c2w control: `camera_controls/<scene_id>/<trajectory_id>/transforms.json`
- Case ID: `<scene_id>__<trajectory_id>`
- 재생성 및 검증: `python build_camera_controls.py`
- 읽기 전용 검증: `python build_camera_controls.py --check`
377개 source × 12개 trajectory로 4,524개 camera-control-only case와
718,188개 control pose를 정의한다. Paired target RGB는 없다.